obtaining parameters of a learning model with a normalization layer

Abstract

Various aspects of this disclosure relate to obtaining parameters of a learning model having normalization layers. An apparatus, such as a network equipment (NE), generates a first set of parameters common to a set of learning models (e.g., machine learning models and / or artificial intelligence (AI) models) and a corresponding second set of parameters for each learning model in the set. The corresponding second set of parameters is the parameters of one or more normalization layers of the learning models in the set. The apparatus transmits a message to another apparatus, such as a user equipment (UE), containing the first set of parameters and at least one second set of parameters. In a variant, the other apparatus requests the second set of parameters and / or both the first set of parameters and the second set of parameters.

Description

Related applications

[0001] This application claims priority to U.S. Provisional Application No. 63 / 574,811, filed April 4, 2024, entitled “Obtaining Parameters for a Learning Model with Normalization Layers,” 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 model techniques for communication. Background Technology

[0003] A wireless communication system may include one or more network communication devices, which may also be referred to as network equipment (NE), supporting wireless communication for one or more user communication devices, which may also 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 a variety of radio access technologies, including third-generation (3G), fourth-generation (4G), fifth-generation (5G), and other suitable radio access technologies other than 5G (e.g., sixth-generation (6G)). Summary of the Invention

[0004] The article “a” preceding an element is unrestricted and should be understood to refer to “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, the word “or,” as used in a list of items (e.g., a list of items beginning with phrases such as “at least one,” “one or more,” or “one or two”) indicates an inclusive list, such that a list of at least one of, for example, 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 referring to a closed set of conditions. For example, without departing from the scope of this disclosure, an example step described as “based on condition A” may be based on both condition A and condition B. 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, the word “group” may comprise one or more elements.

[0005] 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: generate a first set of parameters associated with a corresponding learning model (e.g., a machine learning (ML) model and / or an artificial intelligence (AI) model) in a set of learning models and a corresponding second set of parameters associated with different models in the set of learning models, wherein the corresponding second set of parameters corresponds to one or more corresponding normalization layers of the different models in the set of learning models; and transmit a message to a second device containing the first set of parameters and the second set of parameters in the corresponding second set of parameters.

[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: generate a first set of parameters associated with a corresponding learning model in a set of learning models and a corresponding second set of parameters associated with different models in the set of learning models, wherein the corresponding second set of parameters corresponds to one or more corresponding normalization layers of the different models in the set of learning models; and transmit a message to a second device containing the first set of parameters and the second set of parameters in the corresponding second set of parameters.

[0007] A method for wireless communication, performed by or potentially performed by a first device, is described. The method may include: generating a first set of parameters associated with a corresponding learning model in a set of learning models and a corresponding second set of parameters associated with different models in the set of learning models, wherein the corresponding second set of parameters corresponds to one or more corresponding normalization layers of the different models in the set of learning models; and transmitting a message to a second device containing the first set of parameters and the second set of parameters from the corresponding second set of parameters.

[0008] In some embodiments of the first apparatus, processor, and method described herein, in order to generate a first parameter set associated with the learning model set and a corresponding second parameter set associated with the learning model set, the first apparatus, the processor, and the method may be further configured, capable, or operable to: train a learning model in the learning model set using a subset of data samples from a data sample set to determine the first parameter set and the second parameter set, wherein the trained learning model is associated with the first parameter set and the second parameter set, the second parameter set including scaling and shift parameters associated with one or more normalization layers of the learning model, and the second parameter set including an initial second parameter set of the corresponding second parameter set; and train one or more residual learning models in the learning model set using a corresponding residual data sample subset of the data sample set to determine the corresponding second parameter set, wherein the trained one or more residual learning models are associated with the first parameter set and the corresponding second parameter set. In some embodiments of the first apparatus, processor, and method described herein, the first apparatus, the processor, and the method may be further configured, capable, or operable to determine the data sample set independently of the second apparatus. In some embodiments of the first apparatus, processor, and method described herein, the first apparatus, the processor, and the method may be further configured, enabled, or operable to receive additional messages indicative of the data sample set.

[0009] In some embodiments of the first apparatus, processor, and method described herein, the data sample set comprises a subset of training data samples corresponding to the corresponding learning model, wherein a corresponding subset of training data samples in the set of training data samples is associated with one or more distinct characteristics. In some embodiments of the first apparatus, processor, and method described herein, the one or more distinct characteristics include at least one of the mean of the corresponding subset of training data samples, the standard deviation of the corresponding subset of training data samples, the variance of the corresponding subset of training data samples, or the probability distribution of the corresponding subset of training data samples. In some embodiments of the first apparatus, processor, and method described herein, the set of training data samples is associated with at least one of a corresponding network condition between the first and second apparatuses, a corresponding parameter associated with the first apparatus, a corresponding parameter associated with the second apparatus, or a corresponding channel characteristic associated with a channel established between the first and second apparatuses. In some embodiments of the first apparatus, processor, and method described herein, the data sample set comprises at least one of the following: labeled data samples containing data samples for input to at least one learning model in the learning model set and indications of expected outputs from the at least one learning model; or unlabeled data samples containing data samples for input to the at least one learning model in the learning model set.

[0010] In some embodiments of the first apparatus, processor, and method described herein, the first apparatus, the processor, and the method may be further configured, capable, or operable to generate output from the learning model based on the learning model, which includes the first parameter set and a corresponding second parameter set, based on the learning model to which data is provided as input. In some embodiments of the first apparatus, processor, and method described herein, the first apparatus, the processor, and the method may be further configured, capable, or operable to update the second parameter set associated with the learning model to an additional second parameter set within the corresponding second parameter set. In some embodiments of the first apparatus, processor, and method described herein, the first apparatus, the processor, and the method may be further configured, capable, or operable to determine the corresponding second parameter set based on a corresponding normalization layer associated with the learning model set. In some embodiments of the first apparatus, processor, and method described herein, the corresponding normalization layer corresponds to at least one of a batch normalization layer, a layer normalization layer, a group normalization layer, an example normalization layer, or a root mean square normalization layer. In some embodiments of the first device, processor, and method described herein, the first device, the processor, and the method may be further configured, capable, or operable to transmit, after the message is transmitted, an additional message to the second device containing an additional second set of parameters associated with an additional learning model in the set of learning models, wherein the additional learning model includes the first parameter set and the additional second parameter set.

[0011] In some embodiments of the first device, processor, and method described herein, the first device, the processor, and the method may be further configured, capable, or operable to receive from the second device an additional message instructing the first device to transmit the message, wherein the message is transmitted in response to the additional message. In some embodiments of the first device, processor, and method described herein, the additional message indicates at least one of the first parameter set or the second parameter set. In some embodiments of the first device, processor, and method described herein, in order to transmit the message, the first device, the processor, and the method may be further configured, capable, or operable to determine that one or more conditions associated with the message are satisfied, wherein the one or more conditions are associated with at least one of a corresponding network condition between the first device and the second device, one or more parameters associated with the first device, one or more parameters associated with the second device, or one or more channel characteristics associated with a channel established between the first device and the second device.

[0012] 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: transmit to a second device a first message requesting a first parameter set associated with a corresponding learning model in a learning model set and a second parameter set associated with a learning model in the learning model set, wherein the second parameter set corresponds to one or more normalization layers of the learning model; and receive a second message containing the first parameter set and the second parameter set in response to the first message.

[0013] 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: transmit to a second device a first message requesting a first set of parameters associated with a corresponding learning model in a learning model set and a second set of parameters associated with a learning model in the learning model set, wherein the second set of parameters corresponds to one or more normalization layers of the learning model; and receive a second message containing the first set of parameters and the second set of parameters in response to the first message.

[0014] A method for wireless communication, performed by or potentially performed by a first device, is described. The method may include: transmitting to a second device a first message requesting a first parameter set associated with a corresponding learning model in a learning model set and a second parameter set associated with a learning model in the learning model set, wherein the second parameter set corresponds to one or more normalization layers of the learning model; and receiving a second message containing the first parameter set and the second parameter set in response to the first message.

[0015] In some embodiments of the first apparatus, processor, and method described herein, the second parameter set includes scaling and shifting parameters associated with the one or more normalization layers of the learning model, the second parameter set being determined based on the learning model being trained using a subset of data samples from a data sample set, and the second parameter set including an initial second parameter set associated with a corresponding second parameter set of the learning model set. In some embodiments of the first apparatus, processor, and method described herein, the first apparatus, the processor, and the method may be further configured, capable, or operable to transmit a third message indicating the data sample set to the second apparatus. In some embodiments of the first apparatus, processor, and method described herein, the data sample set includes a subset of training data samples corresponding to the corresponding learning model, wherein a corresponding subset of training data samples in the set of training data samples is associated with one or more distinct characteristics. In some embodiments of the first apparatus, processor, and method described herein, the one or more distinct characteristics include at least one of the mean, standard deviation, variance, or probability distribution of the corresponding training data sample subset. In some embodiments of the first device, processor, and method described herein, the set of training data samples is associated with at least one of the following: a corresponding network condition between the first device and the second device, a corresponding parameter of the first device, a corresponding parameter of the second device, or a corresponding channel characteristic associated with a channel established between the first device and the second device.

[0016] In some embodiments of the first apparatus, processor, and method described herein, the data sample set comprises at least one of the following: labeled data samples containing data samples for input to at least one learning model in the learning model set and an indication of the expected output from the at least one learning model; or unlabeled data samples containing data samples for input to the at least one learning model in the learning model set. In some embodiments of the first apparatus, processor, and method described herein, the first apparatus, the processor, and the method may be further configured, capable, or operable to generate output from the learning model based on providing data as input to the learning model, based on the learning model comprising the first parameter set and the second parameter set. In some embodiments of the first apparatus, processor, and method described herein, the first apparatus, the processor, and the method may be further configured, capable, or operable to: transmit a third message to the second apparatus requesting an additional second parameter set associated with an additional learning model in the learning model set; and receive a fourth message containing the additional second parameter set in response to the third message, wherein the additional learning model comprises the first parameter set and the additional second parameter set.

[0017] In some embodiments of the first apparatus, processor, and method described herein, the one or more normalization layers correspond to at least one of a batch normalization layer, a layer normalization layer, a group normalization layer, an example normalization layer, or a root mean square normalization layer. In some embodiments of the first apparatus, processor, and method described herein, in order to transmit the first message, the first apparatus, the processor, and the method may be further configured, capable of, or operable to determine that one or more conditions associated with the first message are satisfied, wherein the one or more conditions are associated with at least one of a corresponding network condition between the first apparatus and the second apparatus, one or more parameters associated with the first apparatus, one or more parameters associated with the second apparatus, or one or more channel characteristics associated with a channel established between the first apparatus and the second apparatus. Attached Figure Description

[0018] Figure 1 Examples of wireless communication systems according to aspects of this disclosure are described.

[0019] Figure 2 An example of a learning model diagram based on aspects of this disclosure is provided.

[0020] Figure 3 Examples of wireless communication systems according to aspects of this disclosure are described.

[0021] Figure 4 Examples of signaling diagrams according to aspects of this disclosure are provided.

[0022] Figure 5 Examples of UEs based on aspects of this disclosure are described.

[0023] Figure 6 Examples of processors according to aspects of this disclosure are described.

[0024] Figure 7 Examples of NEs based on aspects of this disclosure are described.

[0025] Figure 8 A flowchart illustrating the method performed by the UE according to aspects of this disclosure.

[0026] Figure 9 A flowchart illustrating the method performed by an NE according to aspects of this disclosure. Detailed Implementation

[0027] Wireless communication systems may include one or more devices, such as UEs and NEs, as well as other devices for transmitting and receiving signaling. The devices may implement one or more learning models to perform signaling-related tasks, and these learning models may include AI models and / or ML models. In some variations, the devices may implement learning models for channel state information (CSI) compression (e.g., minimizing CSI complexity and / or encoding CSI into a more compact representation), beam prediction, and positioning procedures, as well as other tasks. In some instances, the NE or other devices train the learning model by determining one or more parameters. For example, the NE selects one or more parameters for the learning model that minimize the loss function of the training dataset. The learning model is then implemented by one or more devices using the determined or selected parameters. However, the devices may input data into a learning model that has different statistical properties (mean, standard deviation, variance, probability distribution, etc.) than the dataset on which the learning model was trained.

[0028] In some cases, the NE can train multiple different learning models on which the learning model makes inferences or predictions, where the domain represents data on scenarios, configurations, network parameters, physical propagation media, or device behavior. The NE can store parameters of the different learning models for transmission to one or more other devices (e.g., the UE or other devices with reduced memory storage capacity compared to the NE). Additionally or alternatively, one or more other devices can receive and store parameters of the different learning models. These other devices can use the corresponding learning models 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, due to the relatively large (e.g., greater than a threshold) number of parameters in the learning models.

[0029] As described herein, to reduce the signaling overhead and processing associated with transmitting and storing parameters of multiple learning models across different domains, the apparatus can generate a parameter set common to multiple learning models and a set of parameters specific to the normalization layer of each learning model. That is, a learning model may contain 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. A normalization layer may contain a relatively small number of parameters (e.g., less than a threshold) compared to a neural layer. Multiple learning models may share the parameters of a neural layer (e.g., have the same parameters), while the normalization layers of different learning models may have different parameters. The apparatus can transmit the parameters of the neural layer and a single set of parameters for the normalization layer of the learning model to another apparatus implementing the learning model in an initial message.

[0030] In some instances, such as for subsequent updates to the learned model, the apparatus may issue different sets of parameters for the normalization layer (e.g., without re-issuing the parameters of the neural layers) because the learned models share the parameters of the neural layers. The apparatus can switch out the parameters of the normalization layer to use different learning models for the corresponding domain. In cases where multiple learning models in different domains share the parameters of the neural layers, the apparatus that issues the parameter set of the normalization layer reduces signaling overhead and processing and storage resources compared to issuing the complete parameter sets of multiple learning models. For example, the number of parameter values ​​in the normalization layer is less than the number of parameter values ​​in the complete parameter set of the learning model, which consumes fewer time-frequency resources, resulting in reduced processing and memory storage.

[0031] In this document, references are made to receiving, transmitting, or conveying data or information, such as signaling communication resources and / or communications transmitted or received between devices. It should be understood that other terms are used interchangeably with conveying, such as signaling, transmitting, receiving, outputting, forwarding, retrieving, obtaining, etc. Similarly, other terms are used interchangeably with transmitting (e.g., conveying, signaling, outputting, forwarding, etc.), and other terms are used interchangeably with receiving (e.g., conveying, retrieving, obtaining, etc.).

[0032] Aspects of this disclosure are described in the context of wireless communication systems.

[0033] Figure 1This section describes an example of a wireless communication system 100 according to 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 network. In some other embodiments, the wireless communication system 100 may be an NR network, such as a 5G network, a 5G-A 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 other than 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).

[0034] One or more NEs 102 may be distributed across a geographical 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., receive signaling, transmit signaling) via a Uu interface.

[0035] NE 102 can provide a geographic coverage area within which it can support services for one or more UEs 104. For example, NE 102 and UE 104 can support wireless communication of signals associated with services (e.g., voice, video, packet data, messaging, broadcasting, etc.) based on one or more radio access technologies. In some embodiments, NE 102 can be mobile, for example, 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 may overlap, but different geographic coverage areas may be associated with different NEs 102.

[0036] One or more UEs 104 may be distributed across a geographical 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 implementations, 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.

[0037] 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.

[0038] NE 102 may support communication with CN 106 or with 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 other network interfaces). In some embodiments, NE 102 may communicate directly with each other. In some other embodiments, NE 102 may communicate with each other or indirectly (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 transmitting entities (which may be referred to as radio heads, smart radio heads, or transmit-receive points (TRPs)).

[0039] CN 106 can support user authentication, access authorization, tracking, connectivity, and other access, routing, or mobility functions. CN 106 can be an evolved packet core (EPC) or a 5G core (5GC), which may include control plane entities (e.g., Mobility Management Entity (MME), Access and Mobility Management Function (AMF)) that manage access and mobility, and user plane entities (e.g., Serving Gateway (S-GW), Packet Data Network (PDN) Gateway (P-GW), or User Plane Function (UPF)) that route packets or interconnect to external networks. In some implementations, the control plane entities may manage non-access stratum (NAS) functions of one or more UEs 104 served by one or more NEs 102 associated with CN 106, such as mobility, authentication, and bearer management (e.g., data bearers, signaling bearers, etc.).

[0040] CN 106 can communicate with the packet data network via one or more backhaul links (e.g., via S1, N2, N6, or other network interfaces). The packet data network may contain an application server. In some implementations, one or more UEs 104 can communicate with the application server. UE 104 can establish a session (e.g., a Protocol Data Unit (PDU) session, etc.) with CN 106 via NE 102. CN 106 can 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 can be an instance of a logical connection between UE 104 and CN 106 (e.g., one or more network functions of CN 106).

[0041] 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 may support different resource structures. For example, NE 102 and UE 104 may support different frame structures. In some embodiments, such as in 4G, NE 102 and UE 104 may support a single frame structure. In some other embodiments, such as in 5G and other suitable radio access technologies, NE 102 and UE 104 may support various frame structures (e.g., multiple frame structures). NE 102 and UE 104 may support various frame structures based on one or more parameter sets.

[0042] 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 regular 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 regular cyclic prefix. A third parameter set (e.g., μ=2) may be associated with a third subcarrier spacing (e.g., 60 kHz) and a regular 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 regular cyclic prefix. A fifth parameter set (e.g., μ=4) may be associated with a fifth subcarrier spacing (e.g., 240 kHz) and a regular cyclic prefix.

[0043] Time intervals for resources (e.g., communication resources) can be organized according to frames (also known as radio frames). Each frame may have a duration, for example, 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 duration, for example, 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.

[0044] Alternatively, 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, the first, second, third, fourth, and fifth parameter sets (e.g., 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 a single 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., OFDM symbols). In some embodiments, the number (e.g., quantity) of time slots in a subframe may depend on the parameter set. For a conventional cyclic prefix, a time slot may contain 14 symbols. For an extended cyclic prefix (e.g., applicable to a 60 kHz subcarrier spacing), a time slot may contain 12 symbols. The relationship between the number of symbols per time slot for the regular cyclic prefix and the extended cyclic prefix, the number of time slots per subframe, and the number of time slots per frame may depend on the parameter set. It should be understood that references 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.

[0045] 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. For example, the wireless communication system 100 may support one or more operating frequency bands, such as frequency range names 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 may perform wireless communication on one or more of the operating frequency bands. In some embodiments, FR1 may 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 can be used by NE 102 and UE 104, as well as other equipment or devices, for short-range, high data rate capabilities.

[0046] 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).

[0047] According to the implementation scheme, one or more of NE 102 and UE 104 can operate to implement various aspects of the technology described with reference to this disclosure. For example, NE 102 (e.g., a base station) transmits a message containing a first set of parameters common to multiple learning models and a second set of parameters specific to the learning model. The first set of parameters may indicate one or more neural layers of the learning model, while the second set of parameters may indicate one or more normalization layers of the learning model. In a variant, the second set of parameters contains a lower number of parameters than the first set of parameters (e.g., less than a threshold number of parameters). NE 102 may transmit a message to UE 104. UE 104 may implement a learning model containing the layers indicated by the first and second set of parameters. Additionally or alternatively, NE 102 may implement a learning model containing the layers indicated by the first and second set of parameters. For example, UE 104 and / or NE 102 may use the learning model to perform CSI compression, beam prediction, and / or positioning procedures.

[0048] In a variant, UE 104 and NE 102 may implement an autoencoder for transmitting and receiving CSI data samples. The autoencoder may be an instance of a neural network architecture or an ML architecture, where data is compressed from its original form to a low-dimensional representation at the encoder, and reconstructed back to its original form at the decoder. The encoder and decoder may be components of the autoencoder, and appropriate learning models may be implemented to perform data compression and reconstruction. In addition to or as an alternative to signal compression, NE 102 and UE 104 may use the autoencoder for future extraction and / or noise reduction. The device may train the autoencoder on a large dataset of wireless signals, and the autoencoder may 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 signals without noise from one or more measurements of a noisy signal, resulting in improved robustness and reliability of the wireless communication system 100 in noisy environments.

[0049] In some instances, such as for signal compression, the encoder receives input data and maps the input data to a low-dimensional representation. The input data may comprise a set of data samples, 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 a channel (e.g., a communication channel) between the transmitter and receiver in the wireless communication system 100. The communication channel is the medium through which data is transmitted from the transmitter to the receiver. In variations, 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 low-dimensional representation may involve using multiple layers of a neural network to transform the input data into a compressed representation of the input data, where the neural network includes one or more learning models (e.g., ML models and / or AI models). The encoder's output is the compressed representation of the input data.

[0050] UE 104 can transmit a compressed representation of the input data to NE 102. NE 102 receives the compressed representation of the input data at a decoder and reconstructs the original input data. The decoder 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 the decoder is a reconstructed version of the input data. For example, NE 102 can reconstruct CSI data to determine one or more characteristics and / or conditions of the channel between UE 104 (e.g., a receiver device) and NE 102 (e.g., a transmitter device). NE 102 can use the CSI data for adaptive modulation and coding, 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.

[0051] In some instances, UE 104 and NE 102 may implement one or more learning models for beam prediction. For example, UE 104 and / or NE 102 may train a learning model to output predictions of transmit and / or receive beams for reliable transmission or reception of signaling. UE 104 and / or NE 102 may input one or more beam measurements into the learning model to obtain predictions of transmit and / or receive beams. Example beam measurements include, but are not limited to, beamforming gain measurements, beamwidth measurements, and signal-to-interference-plus-noise ratio (SINR) measurements, as well as other measurements. Beam prediction may include spatial beam prediction, where UE 104 and / or NE 102 implement a learning model to determine the beams of the available or candidate beam set in the wireless network using a relatively small number of beam measurements (e.g., less than a threshold number of values). For example, the number of values ​​for beam measurements may be less than the number of values ​​for available or candidate beams. Alternatively, beam prediction may include temporal beam prediction, wherein UE 104 and / or NE102 implement a learning model to determine the beam to be used for transmission during one or more subsequent time slots using a relatively small number of historical beam measurements or beam identifiers selected for transmit and / or receive signaling (e.g., less than a threshold number of values).

[0052] In a variant, UE 104 and / or NE 102 implement a learning model to estimate the CSI across the spatiotemporal grid based on pilot and / or reference signals transmitted through a relatively small number of resource elements (e.g., less than a threshold number of values) in the spatiotemporal grid. Additionally or alternatively, UE 104 and / or NE 102 implement a learning model to perform CSI prediction, such that the learning model outputs predictions for one or more future time slots based on the estimated CSI and / or additional information from one or more previous time slots. Additionally or alternatively, UE 104 and / or NE 102 implement a learning model as a receiver. For example, the learning model outputs predictions of transmitted messages, symbols, or bits based on one or more signals received as output from the radio channel.

[0053] In some instances, UE 104 and NE 102 may implement one or more learning models for the location procedure. For example, UE 104 and / or NE 102 implement a learning model to determine the location of a node (e.g., UE 104, NE 102, or any other device in wireless communication system 100) based on the channel impulse response (CIR) experienced by the node. CIR defines how a signal propagates through the channel over time.

[0054] Figure 2 An example of a learning model diagram 200 according to aspects of this disclosure is described. In some instances, the learning model diagram 200 implements aspects of the wireless communication system 100. For example, the learning model diagram 200 may implement aspects of a UE and / or NE, or may be implemented by a UE and / or NE, which may be as described in reference... Figure 1 Examples of UE 104 and NE 102 are described. Learning model diagram 200 illustrates instances of one or more learning models 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.

[0055] In some instances, the learning model 202 may be an instance of an ML model and / or an AI model. For example, one or more learning models 202 may be instances of a neural network with multiple layers. A neural network is a computational model comprising 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. 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 layer 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 that determine the strength of the connections between nodes. During training, the weights are adjusted using input-output pairs from the training dataset with the goal of minimizing a defined loss or error function. Neural networks can learn complex patterns and representations from data, enabling them to perform a wide range of tasks, including classification, regression, clustering, pattern recognition, and sequence generation. Instance neural networks include, but are not limited to, autoencoders (e.g., as referenced). Figure 1 (as described), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Generative Adversarial Network (GAN), Long Short-Term Memory (LSTM) network, or any other type of neural network.

[0056] 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, beam prediction, and / or positioning procedures, as referenced. Figure 1 As described. The UE and / or NE may train one or more learning models 202 on a training dataset to perform CSI compression, beam prediction, and / or localization procedures. 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 the data samples in the dataset. A dataset is called an unlabeled dataset when it contains input data samples but no corresponding output data samples.

[0057] In some instances, the device may implement supervised learning techniques to train the learning model 202. For example, NE 102 may train the learning model 202 to map input data to output labels based on instance 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, enabling the learning model 202 to generalize the mapping to accurately predict new, unseen data. Although 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 procedure. The learning model 202 uses labeled instances while also utilizing the structure of the unlabeled data to improve the performance of the learning model 202.

[0058] In some cases, data samples are set. Represents the labeled training data sample set, where and This represents the input sample and its corresponding label, or equivalently, it represents the input... The expected output samples from learning model 202. It should be noted that... Also known as input-based Prediction and / or inference. In unsupervised learning, the training dataset can be an unlabeled dataset. Learning model 202 is composed of mappings or functions. Definition, where Here, (For example, This represents the optimal set of model parameters learned by model 202 during the training process. The optimal set of model parameters can be the set of parameters that provides the highest performance (e.g., highest accuracy and / or precision, and other performance metrics) for model 202. Determining the optimal model parameter values ​​(e.g., using a dataset) Determine This is referred to as training the learning model 202 or learning from the learning model 202. When Assuming a finite number of values ​​(e.g., when...) ,set When the base number is finite, then the learning model 202 is called a classifier learning model. Assuming continuous values ​​(e.g., when and If (when), then the learning model 202 is called a regression model.

[0059] In some instances, supervised learning and / or training of model 202 involves minimizing a loss function. A loss function (which may additionally or alternatively be called a cost function or objective function) is a mathematical function used to quantify the difference between the predicted output of the learning model 202 and the actual value (e.g., the ground truth). Instance loss functions include, but are not limited to, the mean squared error (MSE) loss function, the cross-entropy loss function, and the hinge loss function, as well as others. In some cases, the optimal parameter set ( This is determined by solving Equation 1:

[0060]

[0061] In some instances, the learning model 202 may be an instance of a deep neural network (DNN). A DNN is a type of artificial neural network (ANN) that contains multiple interconnected layers of neurons or nodes, providing the DNN with a hierarchical representation of data. Unlike shallow neural networks that contain fewer than a threshold number of hidden layers 206 (e.g., one or two), a DNN contains more than a threshold number of hidden layers 206, enabling the DNN to learn complex patterns and relationships in the data. Although supervised learning is described, the device may additionally or alternatively implement unsupervised learning, semi-supervised learning, and / or any other learning techniques to train the learning model 202.

[0062] When implemented (e.g., by a device), the learning model 202 is expected to perform with the same level of accuracy and / or precision as seen during the training and testing phases of the learning model 202 prior to implementation or deployment. However, in real-world wireless networks, the learning model 202 may make predictions and / or inferences based on data with statistical properties different from those on which the learning model 202 was trained. Therefore, the learning model 202 may output incorrect or inaccurate inferences and / or predictions. To improve the accuracy of inferences and / or predictions, the device may train the learning model 202 to generalize to many different data distributions (e.g., distributions different from the training data distribution and those encountered during inference). However, developing a learning model 202 that generalizes to possible domains and outputs the desired performance across those domains can be challenging for the device, especially in the context of wireless networks with different data distribution statistics.

[0063] In some instances, the source domain refers to the set of data samples on which the learning model 202 is trained. For example, when the learning model 202 is training on data samples... During training, This is called the source domain. and In the case of [condition], the joint probability density function (PDF) This represents the source distribution. In the variant, the learning model 202 is trained on more than one source domain or on a dataset containing data samples from more than one source domain, resulting in multiple source domains. , 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, In this case, It is called the target domain, and the 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.

[0064] In some instances, the learning model 202 is used when there is a distribution The learning model 202 is trained on the source domain. Once deployed on the device, the learning model 202 can perform inference and / or prediction in the target domain, where the joint PDF is given by the following: and .when and hour, hint or ,or and Both. Therefore, when the input distribution Change, label distribution Change or conditional distribution When changes occur, the learning model 202 can use the new data distribution for prediction and / or inference. This is especially true when using data from different distributions. When performing inference and / or prediction on input data samples for the target domain, use data with joint PDF. Learning models for source domain data samples and / or training 202 Possibly performing poorly.

[0065] In some instances, the device can use the learning model 202 to perform inference and / or prediction in multiple (e.g., T) target data domains. , Among them, for , To train the learning model 202 to make inferences and / or predictions for T domains, the device can identify the number of possible and / or potential domains on which the learning model 202 will make inferences for the corresponding task (e.g., beam prediction, CSI prediction, and / or CSI-based positioning). For example, for spatial beam prediction, possible and / or potential domains may include domains for indoor users, domains for outdoor users, domains for different codebooks (e.g., those with different beamwidths or radiation patterns), domains for rural coverage areas, domains for urban coverage areas, and domains for dense urban environments affecting physical propagation, as well as domains for different UE orientations, and others. Dataset It may contain data samples from the data domain. For example, data samples may come from the defined scenario and / or configuration of the channel and other relevant network configuration parameters, indicated by k.

[0066] The device can identify different possible domains for a defined task (e.g., beam prediction, CSI prediction, AI-based localization, ML-based localization, AI-based receiver, ML-based receiver, etc.) and can prepare training datasets. The index k iterates through all possible target data domains, scenarios, and / or configurations. The device can use the dataset. , To train and learn model 202 (e.g., DNN). ,in This refers to the number of data domains identified for the task under consideration. The device may store the learning model 202 at a node or device that implements (e.g., uses) the learning model 202 to perform inference and / or prediction in a wireless communication system (e.g., wireless communication system 100). The node and / or device may be a network node and / or device (e.g., a base station or network node responsible for determining location based on measurements reported by edge nodes) or an edge node and / or device, such as a UE. For example, if the learning model 202 is a UE-side model used to help the UE predict the beam to be used by one or more UE measurements (e.g., Reference Signal Received Power (RSRP) values), then the UE may store T models.

[0067] In some cases, training and testing or validating the learning model 202 may utilize a relatively large amount (e.g., greater than a threshold) of computational resources, energy, and power. Alternatively, training and testing the learning model 202 may utilize a relatively large amount (e.g., greater than a threshold) of memory, exceeding the amount of memory, computational resources, energy, and power available to the edge device. Therefore, network nodes (e.g., NEs or centralized facilities without resource constraints) may train the learning model 202. Network nodes may then transmit the trained learning model 202 to other devices where it will be deployed or implemented. Transmitting multiple learning models 202 to the nodes and / or devices where the models will be deployed results in significant costs to network resources (e.g., bandwidth, energy, time, etc.) because the learning model 202 contains a relatively large number (e.g., hundreds or thousands) of parameters transmitted over noisy wireless channels.

[0068] Furthermore, or alternatively, it is difficult to store multiple learning models 202 at an edge device that may have reduced memory or storage capacity compared to the NE (e.g., due to a small form factor). Therefore, the learning models 202 may be stored at the central node and transferred to the edge device. The edge device storing a relatively small number of learning models 202 per task (e.g., less than the threshold number of learning models 202) and receiving other learning models 202 via signaling from the NE or other devices can lead to increased latency associated with the transfer and implementation of the learning models 202. For example, the learning model 202 may contain a relatively large number of parameters (e.g., greater than the threshold number of parameters), and the signaling containing said parameters can consume a relatively significant amount of time (e.g., greater than the threshold duration) to deliver the model from the central node to the edge device. In some instances, learning models 202 with a relatively small number of parameters (e.g., less than the threshold) do not provide prediction and / or inference performance greater than the threshold. Furthermore, techniques such as network pruning, knowledge distillation, model compression (using Kronecker decomposition, model parameter quantization, etc.) to compress and / or reduce the number of numerical values ​​of the parameters of the learning model 202 (e.g., the size of the learning model 202) have not effectively reduced the size of the DNN.

[0069] Conventionally, to train the learning model 202 for inference and / or prediction across different data distributions caused by variations in wireless channel and network parameters, the device may use a hybrid dataset. For example, the device (e.g., NE) may use a training dataset containing data samples from possible data distributions that the learning model 202 can expect to encounter after deployment. However, a learning model 202 trained using multiple data distributions exhibits suboptimal performance across different distributions, while a learning model 202 trained for a specific data distribution will provide better inference performance. Other conventional techniques include model adaptation, where the device trains the learning model 202 for one or two target domains and then adapts and / or adjusts the parameters of the learning model 202 when deploying it for another target domain. However, model adaptation involves retraining the learning model 202 using training samples from the new target domain, which increases the use of computational, energy, power, and memory resources at the device retraining the learning model 202. After deploying the learning model 202 at the device and / or node, the device and / or node may be unable to adapt the learning model 202. If the learning model 202 is deployed on an edge device with a relatively small shape factor (e.g., the size of the edge device is smaller than a threshold size), the edge device may not be able to adapt the model through retraining due to reduced capabilities associated with computational, energy, power, and memory resources at the edge device. In some other conventional techniques, to train the learning model 202 for inference and / or prediction across different data distributions, the device may perform model fine-tuning or refinement, which includes retraining the learning model 202.

[0070] Alternatively, the device may implement multiple learning models 202 with model transfer and switching. However, whenever a learning model 202 is transferred from one node to another (e.g., from NE to UE), transmitting parameters of different learning models 202 to provide edge device model switching results in increased signaling overhead. A learning model 202 with a relatively large number of parameters (e.g., greater than a threshold) may have improved performance compared to a learning model 202 with fewer than a threshold number of parameters, but results in relatively high signaling overhead due to parameter transmission to the device implementing or deploying the learning model 202. Techniques for reducing the size of the learning model 202, such as model compression, can reduce signaling overhead and latency. However, model compression may not effectively reduce signaling overhead and latency to values ​​that satisfy one or more thresholds. In a variant, the device may develop multiple learning models 202 that provide a relatively high level of inference and / or prediction performance on a specific data distribution or target domain (e.g., a specific scenario and / or configuration of a network and / or wireless channel), while ensuring that the parameter differences of the learning models 202 are less than a threshold number, thereby reducing the signaling overhead when transferring different learning models 202 and minimizing the memory used to store the learning models 202.

[0071] The learning model 202 may include one or more normalization layers 210. Normalization or standardization refers to scaling and shifting a data sample set to have zero mean and unit variance. Alternatively, normalization may refer to scaling a data sample set to maintain the values ​​of the data sample set 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 data sample set to have zero mean and unit variance. However, the terms normalization and standardization are used synonymously, where normalization or standardization is defined herein as transforming a data sample set 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, normalization occurs within the neural network (e.g., inside a deep neural network).

[0072] For normalization occurring within a neural network (e.g., learning model 202), the device may feed input data samples to the first layer of the neural network, which may be referred to as input layer 204. If the input data samples have a relatively high dynamic range, where fewer than a threshold number of input samples have values ​​greater than a threshold when 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 the instantaneous gradient values ​​cause the neural network to converge to one or more optimal weights for a relatively long duration (e.g., longer than the threshold duration). In some cases, learning model 202 converges faster during training than if the device normalizes the input weights to have zero mean and unit variance. Hidden layer 206 (e.g., hidden layer 206 after input layer 204) receives input from previous layers. The inputs to the second and larger layers of learning model 202 (e.g., containing the second layer up to 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) are normalized to zero mean and unit variance.

[0073] 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, as well as others. While batch normalization is based on a smaller batch used during training of the learning model 202, layer normalization and group normalization are batch-independent methods, as described herein. Batch normalization reduces the training time of the learning model 202 and stabilizes the model weights for 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 calculated within that portion of the batch of training data samples. That is, batch normalization normalizes the corresponding features within a portion of a batch of training data samples.

[0074] Batch normalization is performed between two or more layers in 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 a batch-normalized (e.g., normalized) version of the input to the subsequent network layer. The learning model 202 has a mean β and a variance γ. 2 It may perform well at conditions below (e.g., rather than at zero mean and unit variance), and the learning model 202 can learn and optimize the mean β and variance γ during training. 2For example, a batch normalization layer can introduce two parameters, β and γ, learned by the learning model 202 during training. In some instances, The features are computed by the layers, 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 ,in and 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 lengths H), i is a three-dimensional vector, where... Batch normalization is defined according to equations 2, 3, and 4:

[0075]

[0076]

[0077]

[0078] In some instances, sets Defined as ,in (and ) represents the sub-indexes of i and j along the C-axis (e.g., the channel axis). In the variant, Representation set The cardinality, and These are relatively small (e.g., less than a threshold) positive constants. For example, in the case of a two-dimensional image as input, pixels with the same channel index are normalized together, meaning batch normalization along... Calculate μ and σ for the axis. Finally. Transformed into ,in and These are learnable and / or trainable parameters, indexed as In the variation, and Multiplication is element-wise multiplication, and Addition is also element-wise addition. Therefore, and With The lengths are the same. For feature i, the parameters This can be referred to as a learnable scaling parameter, and the parameter These can be referred to as learnable shift parameters. Learnable scaling parameters and learnable shift parameters can be referred to as affine parameters of the normalization layer 210. In some cases, the i-th batch of normalization layers has one or more learnable and / or trainable parameters, including... , ,in and The length equals the feature The length.

[0079] During training, features are normalized using the mean and variance computed from the corresponding subsets of the training sample batches across the corresponding subsets of the batches (e.g., mini-batch). During training, the device continuously computes the mean and variance for each batch normalization layer. The expected value, mean, and / or average of subsets spanning the batch, by Indicate, and The expected value, mean, and / or average of subsets spanning the batch, by Indicate. The device will and The values ​​are stored in the i-th batch of normalization layers for inference and / or prediction, for example, when the learning model 202 is implemented at another device and / or the device. In a variant, This indicates the total number of values ​​in a subset of the batch used during training.

[0080] To reduce computation and The computational complexity involved in determining the truth value can be calculated using computationally efficient methods. and The estimate. For example, the device can use the exponential moving average (EMA) method to calculate. and The truth value of , where the device assigns exponentially decreasing weights to past observations. (By ) The EMA representing the mean and the parameters used. small batch variance Calculate based on equations 5 and 6:

[0081]

[0082]

[0083] in It is the total number of subsets of batches (e.g., mini-batches) during training, and It should be noted that and These are 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 a substitute for it, the device may use and ,in yes The estimate (e.g., calculated by a different method than EMA), and yes The estimate is calculated using a different procedure than EMA. Used for estimation and The example 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, while WMA involves assigning different weights to the corresponding data points within a window, providing more recent observations to have a greater impact on the average.

[0084] At the end of training, the j-th batch of normalized layer storage and , ,in It is the number of features at the input of the j-th batch of normalization layers. yes The estimate, and yes The estimate. In the variant, when EMA is used during training to calculate... and When making an estimate, and Therefore, when using EMA, the j-th batch of normalization layer storage... and , ,in This represents the number of features at the input of the j-th batch of normalization layers. In some instances, parameters can be learned without calculating gradients and backpropagation. and Calculated via EMA and In this case, the j-th batch of normalized layers can be learned by simple exponential averaging without calculating gradients and backpropagation. and , .parameter and This is commonly referred to as the batch normalization layer statistic 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 statistics of the j-th batch normalization layer.

[0085] Batch normalization is implemented for learning model 202 to accelerate the training and stabilization of the weights of learning model 202. However, the number of data samples in a subset of a batch can affect the accuracy of the estimated mean and variance of that subset. For example, if the number of data samples in the corresponding subset of a batch is less than a threshold, the error and / or bias of the estimated mean and variance can increase. Similarly, if the number of data samples in the corresponding subset of a batch is greater than a threshold, the error and / or bias of the estimated mean and variance can decrease. In a variant, layer normalization can address the error and / or bias introduced when the number of data samples in the corresponding subset of a batch is less than a threshold.

[0086] In some instances, Let represent the features computed by the layer, and ... ,in and 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 lengths H), i is a three-dimensional vector, where... Layer normalization is defined by equations 2, 3, and 4. For example, the set... Defined as ,in and This represents the sub-indexes of i and j along the B-axis (e.g., the batch axis). In the variant, yes The cardinality, and It is a small (e.g., less than the threshold) positive value.

[0087] Therefore, the device performs layer normalization along (the corresponding sample) The mean μ and standard deviation σ are calculated for the axis. In some instances, Transformed into ,in and These are learnable and / or trainable parameters, indexed as In the variation, and Multiplication is element-wise multiplication, and Addition is also element-wise addition. Therefore, and With The lengths are the same. In some instances, layer normalization normalizes the input samples in nodes within the same layer of the corresponding data samples. Element γ (e.g., all elements γ) may have one or more of the same values, and element β (e.g., all elements β) may have one or more of the same values. For feature i, the parameters This can be referred to as a learnable scaling parameter, and the parameter These can be referred to as learnable shift parameters. These learnable scaling parameters and learnable shift parameters can be referred to as the affine parameters of the normalization layer 210. In a variant, the i-th normalization layer can have one or more learnable and / or trainable parameters. and ,in and The length equals the feature The length of the layer. The device can use layer normalization in any learning model 202 (e.g., regardless of the architecture of the learning model 202).

[0088] Group normalization is another batch-independent normalization technique that a device can use when the number of data samples in the corresponding subset of a batch is less than a threshold. In group normalization, the channel is grouped into... There are 10 groups, each of which Channels, of which This indicates the total number of channels. In the variant, Define the features computed by the layers, where 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 ,in and These represent the batch index, channel index, height index, and width index, respectively. In the case of a one-dimensional input (e.g., each input is a length array H), i is a three-dimensional vector, where... In some instances, group normalization is defined by Equations 2, 3, and 4. For example, the set... Defined as ,in and This represents the sub-indexes of i and j along the B-axis (e.g., the batch axis). and This represents the sub-indexes of i and j along the C-axis (e.g., the channel axis). express The floor, and It is a small (e.g., less than a threshold) positive value. In the variant, Indicates that indices j and i are in the same channel group, where channels are sequentially assigned to the group. For example, channels 1 to... In group 1, and channel arrive In group 2, etc. Therefore, group normalization is calculated along ( ) axis and along ( The mean μ and standard deviation σ of the group of ) channels. For example, for an image, group normalization normalizes all pixels in the group using the same μ and σ.

[0089] In some instances, Transformed into ,in and These are learnable and / or trainable parameters, indexed as In the variation, and Multiplication is element-wise multiplication, and Addition is also element-wise addition. Therefore, and With The lengths are the same. Therefore, the i-th group normalization layer has learnable and / or trainable parameters. and ,in and The length equals the feature The length of . For feature i, the parameter This can be referred to as a learnable scaling parameter, and the parameter These can be referred to as learnable shift parameters. Learnable scaling parameters and learnable shift parameters can be referred to as affine parameters of the normalized layer 210. In some instances, both layer normalization and group normalization are batch-independent because they do not depend on the concepts of batch and batch size. Layer normalization is an instance of group normalization where the number of groups is equal to 1.

[0090] In some instances, the learning model 202 (e.g., DNN) may contain 210 normalization layers, by In some variations, Representing neural networks The set of learnable and / or trainable parameters for the normalized layer 210 in the middle, where yes subsets (e.g.) ),because Representing neural networks 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 batch of normalization layers will introduce... Learnable and / or trainable scaling parameters and A learnable and / or trainable shift parameter In the variation, It is the total number of features at the input of the j-th batch of normalized layers, and the corresponding ( Further includes Scalar parameters (e.g., for) , and ),in yes The length of the feature vector at the input is or equivalently is The number of nodes or neurons in the system.

[0091] At the end of training, the j-th batch of normalized layers storage and , ,in yes The number of features at the input, yes The estimate, and yes The estimate. In the variant, when EMA is used during training to calculate... and When making an estimate, and Therefore, the learnable and / or trainable parameters introduced by the j-th batch of normalization layers also include and , (For example, when using EMA to calculate) and In the case of, they are respectively equivalent to and Therefore, for batch normalization layers... The set of learnable and / or trainable parameters is given by Equation 7:

[0092]

[0093] in yes The total number of features at the input, and yes The length of the feature vector at the input is or equivalently is The number of neurons in the brain.

[0094] When normalization layer 210 is a layer normalization layer, the corresponding layer normalization layer , Each channel and / or feature map has two learnable affine parameters. For example, the learnable affine parameters may include a scaling parameter. and shift parameters Neurons in a normalized layer can have the same scaling and / or weight parameter values ​​and the same bias and / or shift values ​​per channel or feature map. Therefore, when a neural network... The normalization layer 210 in the middle contains a normalization layer. At that time, the set of learnable and / or trainable parameters is given by Equation 8:

[0095]

[0096] When the normalization layer is a group normalization layer At that time, the set of learnable and / or trainable parameters is given by Equation 9:

[0097]

[0098] in yes The total number of features at the input, and yes The length of the feature vector at the input is or equivalently is The number of neurons in the brain.

[0099] In some instances, the device may develop multiple learning models 202 (e.g., DNNs) for a defined task with multiple target domains. The device may train the learning models 202 to make predictions and / or inferences in one or more target domains, which provides reduced usage of memory resources for storing the learning models 202 and reduced signaling overhead for transferring the learning models 202. Instance tasks include, but are not limited to, spatial beam prediction, temporal beam prediction, CSI estimation, CSI prediction, wireless receiver and / or CIR-based localization. For the defined task, the device may use supervised learning to train the learning models 202. Labeled datasets are prepared by collecting data from the field (e.g., by measurements or any collected data operating on a wireless network) or by computer simulations simulating real-world wireless networks. The statistical characteristics of the dataset are determined by multiple factors that are effective during data collection or simulation, such as network configuration (e.g., beamcodebook, multiple-input multiple-output (MIMO) configuration parameters, scheduler type, different types of link adaptation and power control), device behavior (e.g., device mobility, orientation, quality of service (QoS) used to support communication at the device), traffic patterns, and the physical conditions of the propagation medium (e.g., channels with rich multipath versus sparse channels, indoor versus outdoor), etc.

[0100] Although supervised learning is described, the device may additionally or alternatively implement unsupervised or self-supervised learning to train the learning model 202, wherein the training dataset includes an unlabeled dataset. Furthermore or alternatively, the device may implement semi-supervised learning, wherein the learning model 202 is trained using a dataset containing both labeled and unlabeled data samples.

[0101] The defined task can have T target domains. In the variant, The labeled training dataset can represent the target domain k, where as input sample and corresponding tags or expected predictions, inferences and / or outputs The joint distribution over the domain k, where Because domains have different statistical properties, therefore for , The device can use the training dataset. , ( To develop and / or train T neural networks that satisfy a threshold performance level (e.g., threshold prediction accuracy and other performance metrics of the learning model 202) on one of the target domains. , .

[0102] The device can develop learning models 202 that include one or more normalization layers 210. For example, the device can train a normalized layer 210 for one of the target domains. The learning model 202 is represented, where This represents all learnable or trainable parameters or weights of the learning model 202 (e.g., a total of...). The normalization layer 210 can implement any normalization method, such as batch normalization, layer normalization, group normalization, or instance normalization or another normalization method. The normalization layer 210 adjusts and / or changes the statistical distribution of the data samples before using the data samples received from the previous layer as input and output to the next layer.

[0103] The first target domain can be represented as The device uses a labeled training dataset. Supervised learning techniques are used to develop and / or train [learners / students]. By considering the dataset The device can use any of the target domains to develop neural networks. Where k can be assumed to be any value from 1 to T, where, as an example, .

[0104] In order to train the learning model 202 ,exist When used as learnable or trainable parameters, the device can determine the architecture or structure of the learned model 202 (e.g., the number of layers, how the layers are connected, etc.). Instance architectures can include, but are not limited to, CNNs, RNNs, LSTMs, and others. The device can achieve this by minimizing a loss function on the training dataset. To determine one or more values ​​of learnable and / or trainable parameters (e.g., optimal values, including values ​​that maximize the performance of the learned model 202).

[0105] Learning model 202 may include a normalized layer 210 between a pair of hidden layers 206, which may also be referred to as a neural layer. Alternatively, learning model 202 may include a normalized layer 210 preceding the input layer 204 of learning model 202. For example, a four-layer learning model 202 may include one input layer 204, one output layer 208, and two hidden layers 206, without a normalized layer 210. When a normalized layer 210 is introduced into learning model 202, there may be one normalized layer 210 between the input layer 204 and the first hidden layer 206, another normalized layer 210 between the first and second hidden layers 206, and another normalized layer 210 between the second hidden layer 206 and the output layer 208. Therefore, there are three normalized layers 210 in a four-layer learning model 202 (e.g., where there is no normalized layer 210 preceding the input layer 204). In a total of In the learning model 202 with multiple layers, if there is a normalization layer 210 between the corresponding multiple pairs of hidden layers 206, then there will be 210 normalized layers.

[0106] Therefore, the learning model to be developed or trained 202 A normalization layer 210 comprising corresponding pairs of normalization layers (e.g., hidden layer 206), and This represents the learnable parameters of learning model 202. In the variant, The device includes learnable parameters for a normalized layer 210, hidden layer 206, input layer 204, and output layer 208. The normalized layer 210 can be included in any type of learning model 202. Once the network architecture is determined by including the normalized layer 210, the device trains the learning model 202. For example, the device can implement supervised learning techniques to minimize a loss function on a labeled training dataset. The device determines the parameter set that maximizes the performance of the learning model 202 by solving Equation 10 using a selected optimization algorithm (e.g., stochastic gradient descent (SGD) or adaptive moment estimation (ADAM), and others). .

[0107]

[0108] The device can target the remaining Develop and / or train a normalization layer 210 for each target domain. For example, the device can use a learning model 202. This is used to obtain parameters for the corresponding learning model 202 for one or more of the remaining target domains. In some variations, Representation learning model 202 The normalization layer 210 in the middle, and Representation learning model 202 The learnable parameter set of the normalized layer 210 in the model. In some instances, yes subsets (e.g.) ),because Representation learning model 202 The set of learnable and / or trainable parameters. When When the normalization layer 210 is a normalization layer, the normalization layer is... , It has two learnable affine parameters per channel and / or feature map. For example, the learnable affine parameters include a scaling parameter. and shift parameters Neurons in the normalized layer have the same scaling and / or weight parameter values ​​and the same bias and / or shift values ​​for each channel or feature map. Therefore, when learning model 202 Normalization layer 210 in the middle is a normalization layer At that time, the set of learnable and / or trainable parameters is given by Equation 11:

[0109]

[0110] in yes The total number of features / channels at the input.

[0111] when When normalization layer 210 is a batch normalization layer, all normalization layers 210 The set of learnable affine parameters is given by Equation 12:

[0112]

[0113] in yes The characteristics at the input and / or the total number of channels, and yes The length of the feature vector at the input is or equivalently is The number of neurons in the brain.

[0114] At the end of training of model 210, the j-th batch of normalized layer storage and , ,in It is the number of features at the input of the j-th batch of normalization layers. yes The estimate, and yes The estimate. In the variant, when EMA is used during training to calculate... and When making an estimate, and Furthermore, the parameters of the j-th batch of normalized layers can be learned through exponential averaging without requiring gradient computation and backpropagation. and , Therefore, although and These are also learnable parameters in the batch normalized layer, but they are not in the set of learnable affine parameters because... The affine parameters are learned through training. Training includes calculating gradients and backpropagation, while and It can be computed with relatively simple computation based on the forward propagation in learning model 202.

[0115] when Normalization layer 210 in the middle is a group normalization layer At that time, the set of learnable and / or trainable affine parameters is given by Equation 13:

[0116]

[0117] in yes The total number of features at the input, and yes The length of the feature vector at the input is or equivalently is The number of neurons in the brain.

[0118] Learning model 202 can be used for new target domains by changing the parameters of normalization layer 210 without affecting other parameters of learning model 202. Let... and (For example, set) Include set Middle set (Parameters other than those in the parameters). Derived from the parameters The defined DNN is the optimal DNN for target domain 1. The symbol ∪ represents the union operation between two sets. The device computes... This allows the parameters of the DNN for the k-th target domain to be determined by... Given. Therefore, for the remaining A target data domain, which can be used to train and / or develop a separate set or equivalent of a normalized layer 210. The device can use training data from the k-th target domain. ( Training DNN To combine the sets according to Equation 14 All parameters are fixed to previously learned and / or trained values ​​to learn the set. New values ​​for the parameters in:

[0119]

[0120] The device can train and / or develop T learning models 202, wherein the models are based on probability distributions. , The specific target domain of the representation defines the corresponding learning model 202, where for , .

[0121] For example, the device can obtain training datasets for T target domains. , The device can generate A set of parameters: , , This allows the learning model 202 (e.g., DNN) for the target domain k to be derived from... Given, among which The device can select a field from T data fields, defined as the m-th data field, where... For the m-th data field, where ,by As the training dataset, the device minimizes the loss function according to Equation 15. To train a learning model 202 with a normalized layer 210 between multiple pairs of hidden layers (e.g., regular unnormalized neural layers) 206. :

[0122]

[0123] in express Learnable parameters. In some variations, express The normalization layer 210 in the middle, and This indicates all normalized layers 210 A set of learnable and / or trainable parameters. In some instances, In other words, the set Include set Middle set All parameters other than those in the set. In some cases, the set , where parameter set Uniquely determine normalization layer 210 In learning model 202, For the m-th (e.g., The target domain. Therefore, the device determines the parameters. The value and the determination of the normalization layer 210 same.

[0124] For the remaining A target domain (e.g., for) and The device will collect The parameters in are fixed as The previously learned and / or trained values. The device minimizes the training data in the k-th target domain. loss function on (For example, the same loss function used previously) to learn the set for the k-th target domain. The optimal values ​​of the parameters in the text. and To determine In other words, it is determined according to Equation 16. :

[0125]

[0126] Parameter set The normalization layer 210 in the learning model 202 that will be used for the k-th target domain is uniquely determined. Therefore, determining the parameters The value contains the determination of the normalization layer. .

[0127] Figure 3 Examples of a wireless communication system 300 according to aspects of the present invention are described. 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 are described. In some examples, NE 102 may wirelessly communicate with one or more other devices in the wireless communication system 300. For example, UE 104 transmits signaling to NE 102 via uplink communication link 302, which may be as described in reference... Figure 1Examples of described communication links. In some other instances, NE 102 transmits signaling to UE 104 via downlink communication link 304, which may be as described in reference... Figure 1 and 2 An example of a described communication link. Signaling between UE 104 and NE 102 may include control signaling and / or data transmission.

[0128] The apparatus in the wireless communication system 300 can implement one or more learning models (e.g., machine learning models and / or AI models) for various tasks, wherein the learning models can be as referenced. Figure 2 An example of the learning model 202 described. For example, NE 102 and / or UE 104 may implement or deploy one or more learning models for beam prediction, CSI prediction, and as a wireless receiver, as well as other instance tasks. The learning model may contain multiple layers, each containing one or more parameters. Layers may include, but are not limited to, an input layer, at least one hidden layer, at least one normalization layer, and an output layer. In some instances, the parameters of the hidden layers of the learning model may include weights and / or biases that define the relationships between nodes or neurons in the hidden layer. The parameters of the normalization layers of the learning model may include scaling and shifting parameters, as referenced. Figure 2 As described. In some instances, the devices in the wireless communication system 300 can train one or more learning models to obtain or generate parameters for different layers of the learning model. In a variant, NE 102 can train one or more learning models, and NE 102 and UE 104 implement the learning models. For example, NE 102 can generate parameters that define the learning model and can transmit the parameters to UE 104, whereby UE 104 implements the learning model including the parameters. In some other variants, NE 102 and / or UE 104 can both train and implement the learning models.

[0129] Conventionally, NE 102 can develop (e.g., train) multiple learning models for a defined task, wherein the respective learning models have different parameters in the hidden layers and / or may not include a normalization layer. NE 102 can develop learning models for different target domains for the defined task. Different target domains can represent specific scenarios, configurations, network parameters, physical propagation media, or user behavior. That is, the target domain represents data related to scenarios, configurations, network parameters, physical propagation media, and / or user behavior. NE 102 can store the parameters of different learning models for transmission to one or more other devices (e.g., UE 104 or other devices with reduced memory storage capacity compared to NE 102). Additionally or alternatively, one or more other devices can receive and store the parameters of different learning models. Other devices can use the corresponding learning models for different target domains. For example, a device can provide data samples from the target domain as input to at least one learning model. The learning model can provide the device with output for the task. However, transmitting and / or storing the parameters of the learning model for different target domains leads to increased signaling overhead and memory usage as well as inefficient use of time and frequency resources, due to the learning model having a relatively large (e.g., greater than a threshold) number of parameters.

[0130] To reduce signaling overhead, memory usage, and frequency resource consumption, the apparatus in the wireless communication system 300 can implement a learning model with normalized layers. For example, in the initial training process (e.g., for a first learning model), NE 102 can develop or train a complete learning model by obtaining parameters for different layers including hidden layers (e.g., an initial first parameter set) and normalized layers (e.g., an initial second parameter set). For subsequent learning models, NE 102 can develop or train a learning model by obtaining different parameters for the normalized layers (e.g., different second parameter sets) and using the same parameters for the hidden layers as the initial learning model. For example, at 306, NE 102 generates a set of hidden layer parameters common to multiple learning models and different normalized layer parameters for the learning models. That is, multiple learning models share hidden layer parameters and have different parameters for the normalized layers. The number of parameters in the normalized layers can be less than the number of parameters in the hidden layers (e.g., about 1% of the total learning model). Furthermore, or alternatively, the number of parameters in the normalized layers can be less than a threshold (e.g., two parameters per normalized layer). For example, if the learning model has 10 normalized layers, then the NE 102 produces 20 parameter values. Therefore, training multiple learning models by obtaining a single set of parameters for one or more hidden layers and multiple corresponding sets of parameters for one or more normalized layers provides reduced processing and memory usage for storing parameters of different models at the NE 102.

[0131] In some instances, UE 104 may transmit one or more data samples 308 to NE 102. For example, UE 104 may transmit control signaling and / or data transmissions to NE 102 via an uplink communication link 302 containing data samples 308. NE 102 may use data samples 308 to generate parameters for different learning models. The content of data samples 308 may be based on the task to which the learning model will be used. Data samples 308 may contain one or more measurements performed by UE 104, including RSRP measurements, CSI measurements, beamforming measurements, and others. For example, if the task includes CSI data compression, CSI prediction, or CSI estimation, then data samples 308 may contain 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, CSI data contains information describing the characteristics and conditions of the channel between NE 102 and UE 104, such as channel gain data, phase shift data, frequency response data, delay spread data, and Doppler shift data, as well as other information. In some other instances, if the task involves beam prediction, data sample 308 may contain one or more beam measurements, such as beamforming gain measurements, beamwidth measurements, and SINR measurements, as well as other measurements. In some other instances, if the task involves a positioning procedure, data sample 308 may contain one or more CIR measurements, such as timing measurements of signal propagation through the channel between NE 102 and UE 104.

[0132] Alternatively, NE 102 may obtain or otherwise determine data sample 308, for example, by obtaining data sample 308 from a simulation and / or receiving signaling indicating data sample 308 from another device. NE 102 may use data sample 308 as a training dataset to obtain multiple learning models by generating parameters for corresponding learning models. In some instances, data sample 308 may contain labeled and / or unlabeled training data samples, for example, for supervised learning, unsupervised learning, and / or semi-supervised learning. Labeled data samples may contain two components, where the first component is a data sample used as input to the learning model, and the second component is a label of the input sample (e.g., an expected output sample from the learning model). Unlabeled data samples may contain a single component, where the single component is an input sample to the learning model.

[0133] In a variant, NE 102 can partition or group the data sample 308 into subsets of the data sample 308. For example, NE 102 can partition the data sample 308 into subsets of data samples 308 that have different statistical properties (e.g., mean, standard deviation, variance, probability distribution, etc.). That is, NE 102 partitions the data sample 308 such that subsets of the data sample 308 represent the target domain of the task with the defined statistical properties. Alternatively, the data sample 308 can be pre-configured or defined such that the data sample 308 is partitioned into subsets of the data sample 308 before being acquired by NE 102. NE 102 can train a learning model with the same number of numerical values ​​as the subsets of the data sample 308 (e.g., a corresponding learning model for the target domain of the task).

[0134] In some instances, NE 102 may transmit an instruction for a learning model to UE 104 via downlink communication link 304. For example, NE 102 may transmit a hidden layer parameter and normalization layer parameter instruction 310 to UE 104. The hidden layer parameter and normalization layer parameter instruction 310 may include control signaling having one or more fields, the fields having values ​​for parameters of the hidden layer of the learning model and parameters of the normalization layer of the learning model. UE 104 may use the hidden layer parameters and normalization layer parameters to generate a learning model for implementation and / or deployment. For example, UE 104 may use a learning model with hidden layer parameters and a normalization layer to perform a task by providing additional data samples as input to the learning model and receiving output from the learning model.

[0135] In some instances, UE 104 may transmit a request 312 for learning model parameters. This request may include a message in uplink control signaling containing one or more fields instructing NE 102 to transmit hidden layer parameters and normalization layer parameter indications 310. In variations, the request may indicate one or more learning models (e.g., for one or more defined target domains). In some other variations, the request may instruct NE 102 to transmit hidden layer parameters, normalization layer parameters, or both of the learning models. In some instances, UE 104 may transmit multiple requests 312 for learning model parameters. For example, UE 104 may request an initial message from NE 102 indicating both hidden layer parameters and normalization layer parameters. Additionally or alternatively, UE 104 may request one or more subsequent messages from NE 102 indicating normalization layer parameters for one or more learning models, wherein the learning models share a common set of hidden layer parameters and have unique normalization layer parameters. NE 102 may transmit an updated normalized layer parameter indication 314 containing normalized layer parameters of one or more learning models, wherein NE 102 transmits hidden layer parameters to UE 104 before the updated normalized layer parameter indication 314.

[0136] Alternatively, NE 102 can train one or more learning models for implementation at NE 102. NE 102 can select the trained learning model for the task. For example, NE 102 can select a learning model and can provide data samples as input to the learning model. NE 102 can use the output from the learning model for the defined task, as shown in the reference. Figure 1 As described. Although NE 102 is described as training a learning model, UE 104 or another device in wireless communication system 300 may train one or more learning models.

[0137] Figure 4 Examples of signaling diagram 400 according to aspects of this disclosure are illustrated. 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 an example of device 402-a transmitting at least one hidden layer parameter and normalized layer parameter of a learning model to device 402-b. In some cases, device 402-a may be an example of NE 102, and device 402-b may be an example of UE 104, as referenced Figures 1 to 3 The following alternative instances may be implemented, some of which may be executed 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.

[0138] In some instances, at 404, device 402-b sends a request to device 402-a for one or more learning model parameters (e.g., ML model parameters and / or AI model parameters). The request for learning model parameters may include instructions on the learning model and / or instructions on the target domain to which device 402-b will implement the learning model.

[0139] In some cases, at 406, device 402-b transmits one or more data samples to device 402-a. In a variant, device 402-b may transmit the data samples before transmitting a request for learning model parameters. Alternatively, device 402-b may transmit the data samples after transmitting a request for the learning model. Device 402-b may obtain the data samples by performing one or more measurements. For example, the data samples may include RSRP measurements, CSI measurements, and beam measurements, among others. For example, if the task involves CSI data compression, CSI prediction, or CSI estimation, then the data samples may include CSI data obtained by measuring one or more reference signals (e.g., from NE 102). In some cases, the CSI data contains information describing the characteristics and conditions of the communication channel between NE 102 and UE 104, such as channel gain data, phase shift data, frequency response data, delay spread data, and Doppler shift data, among other information. In some other instances, if the task involves beam prediction, the data sample may include one or more beam measurements, such as beamforming gain measurements, beamwidth measurements, SINR measurements, and other measurements. In some other instances, if the task involves a positioning procedure, the data sample may include one or more CIR measurements, such as timing measurements of signal propagation through the communication channel between NE 102 and UE 104. In some other cases, device 402-a determines the data sample independently of device 402-b (e.g., no other indication of signaling or data sample is received from device 402-b). For example, device 402-a may perform measurements to determine the data sample, perform a simulation to generate the data sample, and / or may receive the data sample from another device (e.g., a device other than device 402-b).

[0140] In variations, the data samples may contain subsets of the training data samples. For example, NE 102 may partition the training dataset into subsets of training data samples. In some other instances, NE 102 obtains a training dataset that has already been partitioned into subsets of training data samples. The training dataset contains a subset of training data samples for the corresponding target domain of the task, and NE 102 uses this subset of training data samples to train the learning model. Therefore, NE 102 trains a learning model with the same number of values ​​as the subset of training data samples. NE 102 may partition the training data subset based on the characteristics of the subset. For example, the training data subset may have different characteristics, including but not limited to the mean, standard deviation, variance, or probability distribution of the corresponding training data subset.

[0141] The training data sample subset represents at least one of the following: corresponding network conditions between device 402-a and device 402-b, corresponding parameters associated with device 402-a, corresponding parameters associated with device 402-b, or corresponding channel characteristics of the communication channel established between device 402-a and device 402-b. The training data sample may include network-related measurements or other parameters for communication (e.g., transmitting and receiving signaling) between device 402-a and device 402-b, such as network signal strength, network interference, network channel congestion, network data rate (e.g., throughput), network latency, and other network conditions. Additionally or alternatively, the training data sample may include one or more parameters of the device (e.g., device 402-a and / or device 402-b), such as transmit power for signaling transmitted from the device, antenna gain of the device, modulation scheme used by the device, data rate supported by the device, duplex mode used by the device, and other parameters. Example channel characteristics may include, but are not limited to, attenuation, propagation delay, multipath fading, noise and interference, bandwidth, channel capacity and Doppler shift, and other characteristics.

[0142] In some instances, the data samples comprise labeled data samples and / or unlabeled data samples. Labeled data samples contain data samples used as input to the learning model and indications (e.g., labels) of the expected output from the learning model. Unlabeled data samples contain data samples used as input to the learning model (e.g., without corresponding labels).

[0143] At 408, device 402-a generates a first parameter set common to the learning model set and corresponding second parameter sets for different learning models within the learning model set. That is, device 402-a can perform an initial training process, which includes generating an initial first parameter set and an initial second parameter set for a first learning model. The first parameter set is common to multiple different learning models, such that any subsequent learning model trained by device 402-a shares the first parameter set (e.g., for the hidden layers of the learning models, they have the same first parameter set). The learning models have different second parameter sets, such that any subsequent learning model trained by device 402-a has a different second parameter set (e.g., for one or more normalized layers).

[0144] For example, the first learning model in the learning model set is determined by an initial first parameter set and an initial second parameter set, while the second learning model in the learning model set is determined by an initial first parameter set and another second parameter set different from the initial second parameter set. In some instances, the first parameter set defines the hidden layers of the learning model, and the second parameter set defines the corresponding normalized layers of the learning model. Instance parameters in the first parameter set include, but are not limited to, weights, biases, activation function parameters, and other parameters. Instance parameters in the second parameter set include, but are not limited to, scaling and shifting parameters. Normalized layers may include, but are not limited to, batch normalized layers, layer normalized layers, group normalized layers, instance normalized layers, or root mean square normalized layers.

[0145] In some instances, at 410, device 402-a uses data samples to train a learning model to obtain a first parameter set and a second parameter set. For example, device 402-a uses a subset of data samples from a dataset to train an initial learning model to determine the first parameter set and the second parameter set. In some cases, one or more layers of the initial learning model are defined by the first parameter set and the second parameter set. For example, the second parameter set contains scaling and shift parameters that define one or more normalized layers of the initial learning model. The first parameter set is common to multiple learning models, while the second parameter set is an initial second parameter set specific to the initial learning model.

[0146] In some instances, after training an initial learning model, device 402-a trains one or more residual learning models in the learning model set for the task. Device 402-a can use a corresponding subset of residual data samples from the dataset to determine a second set of parameters for the residual learning models. The learning models in the set (e.g., the initial learning model and the residual learning models) share a common first parameter set and have a unique second parameter set. The first parameter set can contain any number of parameters. Similarly, the second parameter set can contain any number of parameters. The number of parameters in the first parameter set can be greater than the number of parameters in the second parameter set. The number of parameters in the second parameter set can satisfy a threshold (e.g., less than a threshold) such that the signaling overhead for transmitting the second parameter set to another device is relatively small (e.g., less than a threshold).

[0147] At 412, device 402-a transmits a message to device 402-b containing a first parameter set and at least one second parameter set. In a variant, device 402-a transmits the message in response to receiving a request at 404. In some instances, device 402-b transmits the message if one or more conditions are met. For example, device 402-b may determine that corresponding network conditions between device 402-a and device 402-b are met, one or more parameter values ​​of device 402-a are met, one or more parameter values ​​of device 402-b are met, or one or more channel characteristics of the communication channel established between device 402-a and device 402-b are met.

[0148] Alternatively or concurrently, at 414, device 402-a may obtain output from the learning model. That is, device 402-a may select a learning model for a task other than transmitting parameters of one or more learning models to be implemented at device 402-b, or as an alternative. Device 402-a may provide data samples as input to the learning model. The learning model may output information that device 402-a can use for the task. The output may include compressed CSI data, estimated CSI data, predicted CSI data, beampair prediction, positioning information prediction, etc.

[0149] If device 402-a transmits the parameters of the learning model to device 402-b, then at 416, device 402-b can obtain the output from the learning model. For example, device 402-b can use the first parameter set and the second parameter set received at 412 to generate the learning model. Device 402-b can provide data samples as input to the learning model. The learning model can output information that device 402-b can use for the task. The output may include compressed CSI data, estimated CSI data, predicted CSI data, beampair prediction, positioning information prediction, etc.

[0150] In some instances, at 418, device 402-b transmits a request for updated learning model parameters. For example, device 402-b may detect changes in the target domain and request a new learning model for use in the target domain. Changes in the target domain may include, but are not limited to, changes in network conditions between devices 402-a and 402-b, changes in parameters associated with device 402-a, changes in parameters associated with device 402-b, or changes in the channel characteristics of the communication channel established between devices 402-a and 402-b. The request for updated learning model parameters may include an indication of the learning model and / or an indication of detected changes.

[0151] In some cases, at 420, device 402-a transmits an updated second parameter set instruction to device 402-b. Device 402-b uses a learning model with a first parameter set and an updated second parameter set. For example, device 402-b switches the learning model by changing the second parameter set from an existing second parameter set to a new second parameter set. That is, device 402-b switches the normalization layer of the learning model by switching from one second parameter set to another. The updated second parameter set instruction may include the second parameter set of the learning model but not the first parameter set, because the first parameter set is common to the learning model.

[0152] Alternatively, device 402-a can detect changes in the target domain and update the second parameter set of the learning model implemented by device 402-a. For example, device 402-a can switch the second parameter set from an existing parameter set to a new parameter set to update the learning model from the existing learning model to the new learning model.

[0153] Figure 5 An example of a UE 500 according to aspects of this disclosure is described. 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., operatively, communicatively, functionally, electronically, electrically).

[0154] 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 be a processor, digital signal processor (DSP), application-specific integrated circuit (ASIC), or other programmable logic device, or any combination thereof configured to or otherwise support components for performing the functions described in this disclosure.

[0155] 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 to cause UE 500 to perform various functions of this disclosure.

[0156] Memory 504 may comprise 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 media that facilitates the transfer of a computer program from one place to another. Non-transitory storage media may be any available media accessible by a general-purpose or special-purpose computer.

[0157] 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, processor 502 may support wireless communication at UE 500 according to examples disclosed herein. UE 500 (e.g., a first device) may be configured or operable to support components for: transmitting to a second device a first message requesting a first parameter set associated with a corresponding learning model in a learning model set and a second parameter set associated with a learning model in the learning model set, wherein the second parameter set corresponds to one or more normalization layers of the learning model; and receiving a second message containing the first parameter set and the second parameter set in response to the first message.

[0158] Furthermore, UE 500 may be operable to support any or a combination of the following: a second parameter set comprising scaling and shift parameters associated with one or more normalization layers of the learning model; the second parameter set being determined based on training the learning model using a subset of data samples from the data sample set; and the second parameter set comprising an initial second parameter set associated with a corresponding second parameter set of the learning model set. Additionally or alternatively, UE 500 may be operable to support transmitting a third message to a second device indicating a data sample set. Additionally or alternatively, the data sample set comprises a subset of training data samples corresponding to a corresponding learning model, wherein a corresponding subset of training data samples in the set of training data sample samples is associated with one or more distinct characteristics. Additionally or alternatively, the one or more distinct characteristics include at least one of the mean, standard deviation, variance, or probability distribution of the corresponding training data sample subset.

[0159] Additionally or alternatively, the subset of training data samples is associated with at least one of the following: corresponding network conditions between the first and second devices, corresponding parameters of the first device, corresponding parameters of the second device, or corresponding channel characteristics associated with a channel established between the first and second devices. Additionally or alternatively, the data sample set includes at least one of the following: labeled data samples containing data samples for input to at least one learning model in the learning model set and an indication of the expected output from at least one learning model; or unlabeled data samples containing data samples for input to the at least one learning model in the learning model set. Additionally or alternatively, the UE 500 may be operable to support generating output from a learning model based on providing data as input to the learning model, based on the learning model containing a first parameter set and a second parameter set.

[0160] Additionally or alternatively, UE 500 may be operable to support: transmitting a third message to the second device requesting an additional second set of parameters associated with an additional learning model in the learning model set; and receiving a fourth message containing the additional second set of parameters in response to the third message, wherein the additional learning model includes the first set of parameters and the additional second set of parameters. Additionally or alternatively, one or more normalization layers correspond to at least one of a batch normalization layer, a layer normalization layer, a group normalization layer, an example normalization layer, or a root mean square normalization layer. Additionally or alternatively, UE 500 may be operable to support transmitting a first message containing a determination that one or more conditions associated with the first message are satisfied, wherein the one or more conditions are associated with at least one of a corresponding network condition between the first and second devices, one or more parameters associated with the first device, one or more parameters associated with the second device, or one or more channel characteristics associated with a channel established between the first and second devices.

[0161] Alternatively, UE 500 may support: at least one memory (e.g., memory 504); and at least one processor (e.g., processor 502), coupled to at least one memory and configured or operable to cause UE 500 to: transmit to a second device a first message requesting a first parameter set associated with a corresponding learning model in the learning model set and a second parameter set associated with a learning model in the learning model set, wherein the second parameter set corresponds to one or more normalization layers of the learning model; and receive a second message containing the first parameter set and the second parameter set in response to the first message.

[0162] Furthermore, UE 500 may be configured or operable to support any or a combination of the following: a second parameter set includes scaling and shift parameters associated with one or more normalization layers of the learning model; the second parameter set is determined based on training the learning model using a subset of data samples from the data sample set; and the second parameter set includes an initial second parameter set associated with a corresponding second parameter set of the learning model set. Additionally or alternatively, at least one processor is operable to transmit a third message indicating the data sample set to the second device. Additionally or alternatively, the data sample set includes a subset of training data samples corresponding to a corresponding learning model, wherein a corresponding subset of training data samples in the set of training data samples is associated with one or more distinct characteristics. Additionally or alternatively, the one or more distinct characteristics include at least one of the mean, standard deviation, variance, or probability distribution of the corresponding training data sample subset. Alternatively, the subset of training data samples may be associated with at least one of the following: the corresponding network conditions between the first and second devices, the corresponding parameters of the first device, the corresponding parameters of the second device, or the corresponding channel characteristics associated with the channel established between the first and second devices.

[0163] Additionally or alternatively, the data sample set includes at least one of the following: labeled data samples containing data samples for input to at least one learning model in the learning model set and an indication of the expected output from the at least one learning model; or unlabeled data samples containing data samples for input to the at least one learning model in the learning model set. Additionally or alternatively, at least one processor is operable to generate output from a learning model based on providing data as input to the learning model, and based on the learning model containing a first parameter set and a second parameter set. Additionally or alternatively, at least one processor is operable to: transmit a third message to a second device requesting an additional second parameter set associated with an additional learning model in the learning model set; and receive a fourth message containing the additional second parameter set in response to the third message, wherein the additional learning model contains the first parameter set and the additional second parameter set.

[0164] Alternatively or additionally, one or more normalization layers correspond to at least one of batch normalization layer, layer normalization layer, group normalization layer, example normalization layer, or root mean square normalization layer. Alternatively or additionally, in order to transmit the first message, at least one processor is operable to determine that one or more conditions associated with the first message are satisfied, wherein the one or more conditions are associated with at least one of corresponding network conditions between the first and second devices, one or more parameters associated with the first device, one or more parameters associated with the second device, or one or more channel characteristics associated with a channel established between the first and second devices.

[0165] 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.

[0166] 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.

[0167] Receiver chain 510 may be configured to receive signals (e.g., control information, data, packets) over a wireless medium. For example, receiver chain 510 may include one or more antennas for receiving signals over the air or over a wireless medium. 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 the 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 the transmitted data.

[0168] 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, 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 into the air or wireless medium.

[0169] Figure 6An example of a processor 600 according to aspects of this disclosure is described. 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. 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., operatively, communicatively, functionally, electronically, electrically) via one or more interfaces (e.g., buses).

[0170] 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., receive, acquire, retrieve, transmit, output, forward, store, determine, identify, access, write, read) according to the examples described herein. The processor chipset may include one or more cores, one or more caches (e.g., memory native to the processor chipset (e.g., processor 600) or contained within the processor chipset (e.g., processor 600)) or other memory (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).

[0171] Controller 602 can be configured to manage and coordinate various operations of processor 600 (e.g., signaling, receiving, acquiring, retrieving, transmitting, outputting, forwarding, storing, determining, identifying, accessing, writing, and reading) to enable processor 600 to support various operations according to the examples described herein. For example, controller 602 can operate as a control unit of processor 600, generating control signals that manage the operation of various components of processor 600. These control signals include enabling or disabling functional units, selecting data paths, initiating memory accesses, and coordinating operation timing.

[0172] Controller 602 may be configured to fetch (e.g., fetch, 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 operations 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 to enable processor 600 to support various operations according to the examples described herein. Additionally or alternatively, 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.

[0173] Memory 604 may include one or more caches (e.g., memory local to processor 600 or included in processor 600) or other memories, 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., local to processor 600). In some other embodiments, memory 604 may reside outside the processor chipset (e.g., remote from processor 600).

[0174] 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 computer-readable instructions stored in memory 604 to cause processor 600 to perform various functions. For example, processor 600 and / or controller 602 may be coupled to or coupled 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 collectively configured to perform the various functions described herein.

[0175] 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 outside the processor chipset (e.g., processor 600). One or more ALUs 606 may perform one or more calculations 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 circuitry, 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, enabling one or more ALU 606s to handle conditional operations, comparisons, and bitwise operations.

[0176] Processor 600 may support wireless communication according to examples disclosed herein. Processor 600 may be configured or operable to support: at least one controller (e.g., controller 602) coupled to at least one memory (e.g., memory 604) and configured to cause the processor to: transmit to the device a first message requesting a first parameter set associated with a corresponding learning model in a learning model set and a second parameter set associated with a learning model in the learning model set, wherein the second parameter set corresponds to one or more normalization layers of the learning model; and receive a second message containing the first parameter set and the second parameter set in response to the first message.

[0177] Furthermore, the processor 600 may be configured or operable to support any or a combination of the following: a second parameter set comprising scaling and shift parameters associated with one or more normalization layers of the learning model; the second parameter set being determined based on training the learning model using a subset of data samples from the data sample set; and the second parameter set comprising an initial second parameter set associated with a corresponding second parameter set of the learning model set. Additionally or alternatively, at least one controller may be operable to cause the processor to transmit a third message to the device indicating the data sample set. Additionally or alternatively, the data sample set comprises a subset of training data samples corresponding to a corresponding learning model, wherein a corresponding subset of training data samples in the set of training data sample samples is associated with one or more distinct characteristics. Additionally or alternatively, the one or more distinct characteristics include at least one of the mean, standard deviation, variance, or probability distribution of the corresponding training data sample subset.

[0178] Additionally or alternatively, the subset of training data samples is associated with at least one of the following: corresponding network conditions between the processor and the device, corresponding parameters of the processor, corresponding parameters of the device, or corresponding channel characteristics associated with a channel established between the processor and the device. Additionally or alternatively, the data sample set includes at least one of the following: labeled data samples containing data samples for input to at least one learning model in the learning model set and an indication of the expected output from at least one learning model; or unlabeled data samples containing data samples for input to the at least one learning model in the learning model set. Additionally or alternatively, at least one controller is operable to cause the processor to generate output from the learning model based on providing data as input to the learning model, based on the learning model containing a first parameter set and a second parameter set.

[0179] Alternatively or additionally, at least one controller may be operable to cause the processor to: transmit a third message to the device requesting an additional second set of parameters associated with an additional learning model in the learning model set; and receive a fourth message containing the additional second set of parameters in response to the third message, wherein the additional learning model includes the first set of parameters and the additional second set of parameters. Alternatively or additionally, one or more normalization layers correspond to at least one of batch normalization layer, layer normalization layer, group normalization layer, example normalization layer, or root mean square normalization layer. Alternatively or additionally, in order to transmit the first message, at least one controller may be operable to cause the processor to determine that one or more conditions associated with the first message are satisfied, wherein one or more conditions are associated with at least one of a corresponding network condition between the processor and the device, one or more parameters associated with the processor, one or more parameters associated with the device, or one or more channel characteristics associated with a channel established between the processor and the device.

[0180] Figure 7 An example of NE 700 according to aspects of this disclosure is described. 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., operatively, communicatively, functionally, electronically, electrically).

[0181] The 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 be a processor, DSP, ASIC, or other programmable logic device, or any combination thereof configured or otherwise supporting components for performing the functions described in this disclosure.

[0182] 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 to cause NE 700 to perform various functions of this disclosure.

[0183] Memory 704 may comprise 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 medium includes both non-transitory computer storage media and communication media, wherein the communication media includes any media that facilitates the transfer of a computer program from one place to another. Non-transitory storage media may be any available media accessible by a general-purpose or special-purpose computer.

[0184] 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., instructions stored in memory 704 are executed by processor 702). For example, processor 702 may support wireless communication at NE 700 according to examples disclosed herein. NE 700 (e.g., a first device) may be configured or operable to support components for: generating a first set of parameters associated with a corresponding learning model in a learning model set and a corresponding second set of parameters associated with different models in the learning model set, wherein the corresponding second set of parameters corresponds to one or more corresponding normalization layers of different models in the learning model set; and transmitting a message to a second device containing the first set of parameters and the second set of parameters in the corresponding second set of parameters.

[0185] Furthermore, the NE 700 may be configured or operable to support any or a combination of the following: To generate a first parameter set associated with a learning model set and a corresponding second parameter set associated with the learning model set, the method further includes: training a learning model in the learning model set using a subset of data samples from the data sample set to determine the first parameter set and the second parameter set, wherein the trained learning model is associated with the first parameter set and the second parameter set, the second parameter set including scaling and shift parameters associated with one or more normalization layers of the learning model, and the second parameter set including an initial second parameter set for the corresponding second parameter set; and training one or more residual learning models in the learning model set using a corresponding residual data sample subset of the data sample set to determine the corresponding second parameter set, wherein the trained one or more residual learning models are associated with the first parameter set and the corresponding second parameter set. Additionally or alternatively, the NE 700 may be configured or operable to support determining the data sample set independently of the second device. Additionally or alternatively, the NE 700 may be configured or operable to support receiving additional messages indicating the data sample set.

[0186] Alternatively or additionally, the data sample set comprises a subset of training data samples corresponding to a corresponding learning model, wherein the corresponding subset of training data samples in the set of training data samples is associated with one or more different characteristics. Alternatively or additionally, the one or more different characteristics include at least one of the mean, standard deviation, variance, or probability distribution of the corresponding subset of training data samples. Alternatively or additionally, the set of training data samples is associated with at least one of the following: a corresponding network condition between the first and second devices, a corresponding parameter associated with the first device, a corresponding parameter associated with the second device, or a corresponding channel characteristic associated with a channel established between the first and second devices. Alternatively or additionally, the data sample set comprises at least one of the following: labeled data samples containing data samples for input to at least one learning model in the learning model set and an indication of the expected output from at least one learning model; or unlabeled data samples containing data samples for input to the at least one learning model in the learning model set.

[0187] Alternatively, the NE 700 may be configured or operable to support generating output from a learning model based on a learning model that includes a first parameter set and a corresponding second parameter set, based on providing data as input to a learning model set. Alternatively, the NE 700 may be configured or operable to support updating the second parameter set associated with the learning model to an additional second parameter set within the corresponding second parameter set. Alternatively, the NE 700 may be configured or operable to support determining a corresponding second parameter set based on a corresponding normalization layer associated with the learning model set. Alternatively, the corresponding normalization layer corresponds to at least one of a batch normalization layer, a layer normalization layer, a group normalization layer, an example normalization layer, or a root mean square normalization layer. Alternatively, the NE 700 may be configured or operable to support transmitting an additional message to a second device after a message is transmitted, containing an additional second parameter set associated with an additional learning model in the learning model set, wherein the additional learning model includes the first parameter set and the additional second parameter set.

[0188] Alternatively or additionally, the NE 700 may be configured or operable to support receiving additional messages from the second device that instruct the first device to transmit a message, wherein the message is transmitted in response to the additional message. Alternatively or additionally, the additional message indicates at least one of a first parameter set or a second parameter set. Alternatively or additionally, in order to transmit a message, the NE 700 may be configured or operable to support determining that one or more conditions associated with the message are satisfied, wherein the one or more conditions are associated with at least one of the following: corresponding network conditions between the first and second devices, one or more parameters associated with the first device, one or more parameters associated with the second device, or one or more channel characteristics associated with a channel established between the first and second devices.

[0189] Alternatively, NE 700 may support: at least one memory (e.g., memory 704); and at least one processor (e.g., processor 702), coupled to at least one memory and configured or operable to cause NE to: generate a first set of parameters associated with a corresponding learning model in the learning model set and a corresponding second set of parameters associated with a different model in the learning model set, wherein the corresponding second set of parameters corresponds to one or more corresponding normalization layers of the different models in the learning model set; and transmit a message to a second device containing the first set of parameters and the second set of parameters in the corresponding second set of parameters.

[0190] Furthermore, the NE 700 may be operable or configured to support any or a combination of the following: To generate a first parameter set associated with a learning model set and a corresponding second parameter set associated with the learning model set, at least one processor may operate the NE 700 to: train a learning model in the learning model set using a subset of data samples from the data sample set to determine the first and second parameter sets, wherein the trained learning model is associated with the first and second parameter sets, the second parameter set containing scaling and shift parameters associated with one or more normalization layers of the learning model, and the second parameter set containing an initial second parameter set for the corresponding second parameter set; and train one or more residual learning models in the learning model set using a corresponding residual data sample subset of the data sample set to determine the corresponding second parameter set, wherein the trained one or more residual learning models are associated with the first and corresponding second parameter sets. Additionally or alternatively, at least one processor may operate the NE 700 to determine the data sample set independently of the second means. Additionally or alternatively, at least one processor may operate the NE 700 to receive additional messages indicating the data sample set.

[0191] Alternatively or additionally, the data sample set comprises a subset of training data samples corresponding to a corresponding learning model, wherein the corresponding subset of training data samples in the set of training data samples is associated with one or more different characteristics. Alternatively or additionally, the one or more different characteristics include at least one of the mean, standard deviation, variance, or probability distribution of the corresponding subset of training data samples. Alternatively or additionally, the set of training data samples is associated with at least one of the following: a corresponding network condition between the first and second devices, a corresponding parameter associated with the first device, a corresponding parameter associated with the second device, or a corresponding channel characteristic associated with a channel established between the first and second devices. Alternatively or additionally, the data sample set comprises at least one of the following: labeled data samples containing data samples for input to at least one learning model in the learning model set and an indication of the expected output from at least one learning model; or unlabeled data samples containing data samples for input to the at least one learning model in the learning model set.

[0192] Alternatively, at least one processor is operable to cause the NE 700 to generate an output from a learning model based on a learning model that provides data as input to a learning model set, and based on a learning model containing a first parameter set and a corresponding second parameter set. Alternatively, at least one processor is operable to cause the NE 700 to update the second parameter set associated with the learning model to an additional second parameter set within the corresponding second parameter set. Alternatively, at least one processor is operable to cause the NE 700 to determine a corresponding second parameter set based on a corresponding normalization layer associated with the learning model set. Alternatively, the corresponding normalization layer corresponds to at least one of a batch normalization layer, a layer normalization layer, a group normalization layer, an example normalization layer, or a root mean square normalization layer. Alternatively, at least one processor is operable to cause the NE 700 to transmit, after a message is transmitted, an additional message to a second device containing an additional second parameter set associated with an additional learning model in the learning model set, wherein the additional learning model contains the first parameter set and the additional second parameter set.

[0193] Alternatively or additionally, at least one processor may be operable to cause NE 700 to receive from the second device an additional message instructing the first device to transmit a message, wherein the message is transmitted in response to the additional message. Alternatively or additionally, the additional message instructs at least one of a first parameter set or a second parameter set. Alternatively or additionally, in order to transmit the message, at least one processor may be operable to cause NE 700 to determine that one or more conditions associated with the message are satisfied, wherein the one or more conditions are associated with at least one of a corresponding network condition between the first and second devices, one or more parameters associated with the first device, one or more parameters associated with the second device, or one or more channel characteristics associated with a channel established between the first and second devices.

[0194] 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.

[0195] 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.

[0196] Receiver chain 710 may be configured to receive signals (e.g., control information, data, packets) over a wireless medium. For example, receiver chain 710 may include one or more antennas for receiving signals over the air or over a wireless medium. 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 the 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 the transmitted data.

[0197] 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, 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 into the air or wireless medium.

[0198] Figure 8 A flowchart illustrating method 800 according to an aspect of this disclosure is provided. 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 method described herein describes possible implementations, and the operation and steps can be rearranged or otherwise modified, and other implementations are possible.

[0199] At 802, the method may include transmitting to a second device a first message requesting a first parameter set associated with a corresponding learning model in the learning model set and a second parameter set associated with a learning model in the learning model set, wherein the second parameter set corresponds to one or more normalization layers of the 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 as described in references... Figure 5 The UE execution described.

[0200] At 804, the method may include receiving a second message comprising a first parameter set and a second parameter set in response to the first message. The operation of 804 may be performed according to the examples described herein. In some embodiments, aspects of the operation of 804 may be as described in references... Figure 5 The UE execution described.

[0201] Figure 9 A flowchart illustrating method 900 according to an aspect of this disclosure is provided. The operation of the method may 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 method described herein describes possible implementations, and the operation and steps may be rearranged or otherwise modified, and other implementations are possible.

[0202] In 902, the method may include generating a first set of parameters associated with a corresponding learning model in the learning model set and a corresponding second set of parameters associated with different models in the learning model set, wherein the corresponding second set of parameters corresponds to one or more corresponding normalization layers of different models in the learning model set. The operation of 902 may be performed according to the examples described herein. In some embodiments, aspects of the operation of 902 may be provided by reference to [reference needed]. Figure 7 The described NE execution.

[0203] At 904, the method may include transmitting a message to a second device containing a first parameter set and a corresponding second parameter set. The operation of 904 may be performed according to the examples described herein. In some embodiments, aspects of the operation of 904 may be as described in references... Figure 7 The described NE execution.

[0204] 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 readily 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 should be given the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A first device for wireless communication, comprising: At least one memory; and At least one processor, coupled to the at least one memory and operable to cause the first device to: Generate a first set of parameters associated with a corresponding learning model among a plurality of learning models and a corresponding second set of parameters associated with a different one among the plurality of learning models, wherein the corresponding second set of parameters corresponds to one or more corresponding normalization layers of the different one among the plurality of learning models; and A message including the first parameter set and the second parameter set in the corresponding second parameter set is transmitted to the second device.

2. The first apparatus of claim 1, wherein, in order to generate the first parameter set associated with the plurality of learning models and the corresponding second parameter set associated with the plurality of learning models, the at least one processor is operable to cause the first apparatus to: The learning model among the multiple learning models is trained using a subset of data samples from the data sample set to determine the first parameter set and the second parameter set, wherein: The trained learning model is associated with the first parameter set and the second parameter set; The second parameter set includes scaling and shift parameters associated with one or more normalization layers of the learning model; and The second parameter set includes the initial second parameter set of the corresponding second parameter set; and One or more of the plurality of learning models are trained using a corresponding subset of the remaining data samples of the data sample set to determine the corresponding second parameter set, wherein the trained one or more remaining learning models are associated with the first parameter set and the corresponding second parameter set.

3. The first apparatus of claim 2, wherein the at least one processor is further operable to enable the first apparatus to determine the data sample set independently of the second apparatus.

4. The first apparatus of claim 2, wherein the at least one processor is further operable to cause the first apparatus to receive an additional message indicating the data sample set.

5. The first apparatus according to claim 2, wherein the data sample set includes a plurality of training data sample subsets corresponding to the corresponding learning model, and wherein a corresponding training data sample subset in the plurality of training data sample subsets is associated with one or more different features.

6. The first apparatus of claim 5, wherein the one or more different characteristics include at least one of the mean of the corresponding training data sample subset, the standard deviation of the corresponding training data sample subset, the variance of the corresponding training data sample subset, or the probability distribution of the corresponding training data sample subset.

7. The first apparatus of claim 5, wherein the plurality of training data sample subsets are associated with at least one of a corresponding network condition between the first apparatus and the second apparatus, a corresponding parameter associated with the first apparatus, a corresponding parameter associated with the second apparatus, or a corresponding channel characteristic associated with a channel established between the first apparatus and the second apparatus.

8. The first apparatus of claim 2, wherein the data sample set comprises at least one of the following: The labeled data samples include data samples for input to at least one of the plurality of learning models and indications of the expected output from the at least one learning model; or Unlabeled data samples, which include data samples used as input to at least one of the plurality of learning models.

9. The first apparatus of claim 1, wherein the at least one processor is further operable to cause the first apparatus to: At least partially based on a learning model that provides data as input to the plurality of learning models, and at least partially based on the learning model that includes the first parameter set and a corresponding second parameter set to generate output from the learning model; and The second parameter set associated with the learning model is updated to an additional second parameter set of the corresponding second parameter set.

10. The first apparatus of claim 1, wherein the at least one processor is further operable to cause the first apparatus to determine the corresponding second parameter set at least in part based on a corresponding normalization layer associated with the plurality of learning models, wherein the corresponding normalization layer corresponds to at least one of a batch normalization layer, a layer normalization layer, a group normalization layer, an example normalization layer, or a root mean square normalization layer.

11. The first apparatus of claim 1, wherein the at least one processor is further operable to cause the first apparatus to transmit, after the message is transmitted, an additional message to the second apparatus comprising an additional second set of parameters associated with an additional learning model among the plurality of learning models, and wherein the additional learning model comprises the first parameter set and the additional second parameter set.

12. The first apparatus of claim 1, wherein the at least one processor is further operable to cause the first apparatus to receive from the second apparatus an additional message instructing the first apparatus to transmit the message, wherein the message is transmitted in response to the additional message, and wherein the additional message instructs at least one of the first parameter set or the second parameter set.

13. The first apparatus of claim 1, wherein, in order to transmit the message, the at least one processor is operable to cause the first apparatus to determine that one or more conditions associated with the message are satisfied, and wherein the one or more conditions are associated with at least one of a corresponding network condition between the first apparatus and the second apparatus, one or more parameters associated with the first apparatus, one or more parameters associated with the second apparatus, or one or more channel characteristics associated with a channel established between the first apparatus and the second apparatus.

14. A method performed by a first device, the method comprising: Generate a first set of parameters associated with a corresponding learning model among a plurality of learning models and a corresponding second set of parameters associated with a different one among the plurality of learning models, wherein the corresponding second set of parameters corresponds to one or more corresponding normalization layers of the different one among the plurality of learning models; and A message including the first parameter set and the second parameter set in the corresponding second parameter set is transmitted to the second device.

15. A first device for wireless communication, comprising: At least one memory; and At least one processor, coupled to the at least one memory and operable to cause the first device to: A first message is sent to a second device requesting a first parameter set associated with a corresponding learning model among a plurality of learning models and a second parameter set associated with a learning model among the plurality of learning models, wherein the second parameter set corresponds to one or more normalization layers of the learning model; and In response to the first message, a second message including the first parameter set and the second parameter set is received.

16. The first apparatus according to claim 15, wherein: The second parameter set includes scaling and shifting parameters associated with the one or more normalization layers of the learning model; The second parameter set is determined at least in part based on training the learning model using a subset of data samples from the data sample set; and The second parameter set includes an initial second parameter set associated with the corresponding second parameter sets of the plurality of learning models.

17. The first apparatus of claim 15, wherein the at least one processor is further operable to cause the first apparatus to generate an output from the learning model based at least in part on providing data as input to the learning model, and at least in part on the learning model including the first parameter set and the second parameter set.

18. The first apparatus of claim 15, wherein the at least one processor is further operable to cause the first apparatus to: A third message is sent to the second device requesting an additional second set of parameters associated with an additional learning model among the plurality of learning models; and In response to the third message, a fourth message including the additional second parameter set is received, wherein the additional learning model includes the first parameter set and the additional second parameter set.

19. The first apparatus of claim 15, wherein, in order to transmit the first message, the at least one processor is operable to cause the first apparatus to determine that one or more conditions associated with the first message are satisfied, and wherein the one or more conditions are associated with at least one of a corresponding network condition between the first apparatus and the second apparatus, one or more parameters associated with the first apparatus, one or more parameters associated with the second apparatus, or one or more channel characteristics associated with a channel established between the first apparatus and the second apparatus.

20. A method performed by a first device, the method comprising: A first message is sent to a second device requesting a first parameter set associated with a corresponding learning model among a plurality of learning models and a second parameter set associated with a learning model among the plurality of learning models, wherein the second parameter set corresponds to one or more normalization layers of the learning model; and In response to the first message, a second message including the first parameter set and the second parameter set is received.

Images