End-to-end machine learning for wireless networks

Abstract

The invention relates to end-to-end machine learning for wireless networks. Techniques and apparatus for generating end-to-end machine learning configurations for wireless networks are described. An end-to-end machine learning controller determines an end-to-end machine learning configuration for processing information exchanged over an end-to-end communication in a wireless network. The end-to-end machine learning controller obtains capabilities of one or more devices utilized in the end-to-end communication. The end-to-end machine learning controller determines an end-to-end machine learning configuration for processing information exchanged over the end-to-end communication, and directs one or more devices to process information exchanged over the end-to-end communication by forming one or more deep neural networks based on the end-to-end machine learning configuration.

Description

Description of the case This application is a divisional application of Chinese invention patent application No. 201980100522.X, filed on October 28, 2019. Technical Field The present disclosure generally relates to end-to-end machine learning for wireless networks. Background Art

[0001] The evolution of wireless communication systems often stems from the demand for data throughput. As an example, as more and more devices gain access to wireless communication systems, the demand for data increases. Evolved devices also execute data-intensive applications that utilize more data than traditional applications, such as data-intensive streaming video applications, data-intensive social media applications, data-intensive audio services, etc. This growing demand can sometimes exhaust the data resources of the wireless communication system. Therefore, in order to accommodate the increased data usage, the evolved wireless communication systems utilize increasingly complex architectures to provide more data throughput relative to traditional wireless communication systems.

[0002] As an example, fifth generation (5G) standards and technologies use higher frequency ranges (e.g., 5G millimeter wave (mmW) technology) such as bands above 6 gigahertz (GHz) to transmit data to increase data capacity. However, using these higher frequency ranges to transmit and recover information presents challenges. To illustrate, high frequency signals are more susceptible to multipath fading, scattering, atmospheric absorption, diffraction, interference, etc. relative to low frequency signals. These signal distortions often result in errors when the information is recovered at the receiver. As another example, hardware capable of transmitting, receiving, routing and / or otherwise using these higher frequencies can be complex and expensive, which increases processing costs in wireless network devices. Summary of the invention

[0003] This document describes techniques and apparatus for end-to-end (E2E) machine learning for wireless systems. Various aspects describe an end-to-end machine learning controller that obtains capabilities of devices associated with end-to-end communications in a wireless network. In some implementations, the end-to-end machine learning controller obtains machine learning (ML) capabilities of at least one device. In response to obtaining the capabilities, the end-to-end machine learning controller determines an end-to-end machine learning configuration (E2E ML configuration) based on the capabilities of the devices, and communicates the end-to-end machine learning configuration to one or more devices. Alternatively or additionally, the end-to-end machine learning controller directs each of the devices to form a corresponding deep neural network based on the end-to-end machine learning configuration, and uses the corresponding deep neural network to process the end-to-end communication.

[0004] One or more aspects describe a user equipment (UE) that communicates one or more capabilities supported by the user equipment to an end-to-end machine learning controller, such as implemented at a network entity. The UE receives a neural network formation configuration based on the end-to-end machine learning configuration for processing information exchanged through the end-to-end communication. The UE uses the neural network formation configuration to form a deep neural network, and then uses the deep neural network to process the information exchanged through the end-to-end communication.

[0005] Details of one or more implementations of end-to-end machine learning for wireless networks are set forth in the accompanying drawings and the following description. Other features and advantages will be apparent from the description and drawings, and from the claims. This summary is provided to introduce subject matter that will be further described in the detailed implementations and drawings. Therefore, this summary should not be considered to describe essential features, nor should it be used to limit the scope of the claimed subject matter. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] The following describes details of one or more aspects of end-to-end machine learning for wireless networks. The same reference numerals are used in different instances in the specification and drawings to indicate similar elements: Figure 1 Illustrated is an example environment in which various aspects of end-to-end machine learning for wireless networks can be implemented. Figure 2 An example device diagram illustrates devices that can implement various aspects of end-to-end machine learning for wireless networks. Figure 3 An example device diagram illustrates devices that can implement various aspects of end-to-end machine learning for wireless networks. Figure 4 Illustrated are example machine learning modules that can implement various aspects of end-to-end machine learning for wireless networks. Figure 5 An example block diagram of a processing chain used by a device to process communications transmitted over a wireless communication system is illustrated. Figure 6 An example operating environment for utilizing multiple deep neural networks in a wireless communication system is illustrated. Figure 7 An example transaction graph between various devices for configuring a neural network using a neural network formation configuration is illustrated. Figure 8 An example transaction graph between various devices for updating a neural network using a neural network formation configuration is illustrated. Figure 9-1 and Figure 9-2 An example transaction graph between various devices for configuring a neural network using a neural network formation configuration is illustrated. Fig.10 An example method for configuring a neural network to process communications transmitted over a wireless communication system is illustrated. Fig.11 An example method for forming a neural network based on a neural network formation configuration is illustrated. Fig.12 An example of generating multiple neural network formation configurations is illustrated. Fig.13 An example transaction diagram between various devices for communicating a neural network formation configuration is illustrated. Fig.14 An example transaction graph between various network entities for communicating a neural network formation configuration is illustrated. Fig.15 An example transaction graph between various network entities for communicating a neural network formation configuration is illustrated. Fig.16 An example method for communicating a neural network formation configuration over a wireless communication system is illustrated. Fig.17 An example method for communicating a neural network formation configuration over a wireless communication system is illustrated. Fig.18 An example environment illustrating example changes between QoS flows is illustrated. Fig.19 Illustrated is an example environment in which various aspects of end-to-end machine learning for wireless networks can be implemented. Fig. 20 Illustrated is an example environment in which various aspects of end-to-end machine learning for wireless networks can be implemented. Fig.21 Illustrated are example transaction graphs between various devices in accordance with various implementations of end-to-end machine learning for wireless networks. Fig. 22 Illustrated are example transaction graphs between various devices in accordance with various implementations of end-to-end machine learning for wireless networks. Fig.23 Illustrated are example transaction graphs between various devices in accordance with various implementations of end-to-end machine learning for wireless networks. Fig.24 Illustrated are example transaction graphs between various devices in accordance with various implementations of end-to-end machine learning for wireless networks. Fig.25 An example approach for end-to-end machine learning for wireless networks is illustrated. Fig.26 An example approach for end-to-end machine learning for wireless networks is illustrated. DETAILED DESCRIPTION

[0007] In traditional wireless communication systems, the transmitter and receiver processing chains include complex functions. For example, a channel estimation block in the processing chain estimates or predicts how the transmission environment distorts the signal propagating through the transmission environment. As another example, the channel equalizer block inverts the distortion identified from the signal by the channel estimation block. These complex functions often become more complex when processing higher frequency ranges (such as 5G mmW signals in the 6 GHz range or near it). For example, the transmission environment adds more distortion to the higher frequency ranges relative to the lower frequency ranges, thus making information recovery more complex. In addition, hardware capable of processing and routing higher frequency ranges often adds increased costs and complex physical constraints.

[0008] DNNs provide solutions to complex processing, such as complex functions used in wireless communication systems. By training DNNs on transmitter and / or receiver processing chain operations, DNNs can replace traditional complex functions in various ways, for example, by replacing some or all traditional processing blocks used in end-to-end processing of wireless communication signals, replacing individual processing chain blocks, etc. Dynamic reconfiguration of DNNs, such as by modifying various parameter configurations (e.g., coefficients, layer connections, kernel size), also provides the ability to adapt to changing operating conditions.

[0009] This document describes aspects of end-to-end machine learning for wireless networks that allow the system to process communications and dynamically reconfigure DNNs used in E2E communications as operating conditions change. The aspects may be implemented by network entities (e.g., base stations, core network servers, user equipment) operating in a wireless communication system.

[0010] In an implementation, a network entity obtains capabilities of devices associated with E2E communication in a wireless network. In some implementations, the network entity requests capabilities, such as by requesting UE capabilities, machine learning capabilities (ML capabilities), base station (BS) capabilities, etc. from one or more devices participating in the E2E communication. In response to obtaining the capabilities, the network entity determines an end-to-end machine learning configuration (E2E ML configuration) based on the received capabilities. After determining the E2E ML configuration, the network entity communicates the E2E ML configuration to the device, such as by communicating a first neural network (NN) format configuration to the first device, a second NN format configuration to the second device, etc. The network entity then directs the device participating in the E2E communication to form a corresponding deep neural network (e.g., a first DNN formed using the first NN format configuration, a second DNN formed using the second NN format configuration) based on the E2E ML configuration, and uses the corresponding DNN to process the E2E communication.

[0011] In an implementation, the UE transmits one or more capabilities supported by the user equipment to a network entity. For example, the UE transmits processing capabilities, memory capabilities, ML support capabilities, etc. Thereafter, the UE receives a NN information configuration associated with the E2E ML configuration. The UE then forms a DNN using the NN formation configuration, and uses the DNN to process UE-side communications (e.g., the UE portion of the E2E communication).

[0012] Determining an E2E ML configuration (where the E2E ML configuration may be partitioned among multiple devices) enables the network entity to select an ML configuration that improves communication exchanges in the system. For example, in some implementations, the network entity partitions the E2E ML configuration based on device capabilities such that the network entity directs devices with less processing resources and / or memory to form a DNN with less processing (e.g., less data, less memory, less CPU cycles, fewer layers) relative to devices with more processing resources and / or memory. In other words, the network entity determines an E2E ML configuration that improves the overall performance of E2E communications exchanged between devices (e.g., lower bit errors, improved signal quality, improved latency), and partitions the E2E ML configuration based on device resources to allocate and / or appropriate processing in an efficient manner.

[0013] The phrases "transmitted via," "communications exchanged," and "communications associated with" include generating communications to be transmitted via a wireless communication system (e.g., processing communications prior to transmission) and / or processing communications received via a wireless communication system. Thus, "processing communications transmitted via a wireless communication system," "communications exchanged via a wireless communication system," and "communications associated with a wireless communication system" include generating transmissions (e.g., processing prior to transmission), processing received transmissions, or any combination thereof. Example Environment

[0014] Figure 1An example environment 100 is illustrated including a user equipment 110 (UE 110) capable of communicating with a base station 120 (illustrated as base stations 121 and 122) via one or more wireless communication links 130 (wireless links 130) illustrated as wireless links 131 and 132. For simplicity, the user equipment 110 is implemented as a smart phone, but may be implemented as any suitable computing or electronic device, such as a mobile communication device, a modem, a cellular phone, a gaming device, a navigation device, a media device, a laptop, a desktop computer, a tablet computer, a smart home appliance, or a vehicle-based communication system or an Internet of Things (IoT) such as a sensor or actuator. The base station 120 (e.g., an evolved universal terrestrial radio access network Node B, E-UTRAN Node B, evolved Node B, eNodeB, eNB, next generation Node B, gNode B, gNB, ng-eNB, etc.) may be implemented in a macro cell, a micro cell, a small cell, a pico cell, etc., or any combination thereof.

[0015] The base station 120 communicates with the user equipment 110 using radio links 131 and 132, which may be implemented as any suitable type of radio links. The radio links 131 and 132 include control and data communications, such as downlinks of data and control information communicated from the base station 120 to the user equipment 110, uplinks of other data and control information communicated from the user equipment 110 to the base station 120, or both. The radio link 130 may include one or more radio links (e.g., radio links) or bearers implemented using any suitable communication protocol or standard or a combination of communication protocols or standards such as the following: Third Generation Partnership Project Long Term Evolution (3GPP LTE), Fifth Generation New Radio (5G NR), etc. Multiple radio links 130 may be aggregated in carrier aggregation to provide higher data rates for the UE 110. Multiple radio links 130 from multiple base stations 120 may be configured for coordinated multi-point (CoMP) communication with the UE 110.

[0016] The base stations 120 are collectively a radio access network 140 (e.g., a RAN, an evolved universal terrestrial radio access network, an E-UTRAN, a 5G NR RAN, or an NR RAN). The base stations 121 and 122 in the RAN 140 are connected to a core network 150. The base stations 121 and 122 are connected to the core network 150 at 102 and 104, respectively, via an NG2 interface for control plane signaling, and an NG3 interface for user plane data communication when connected to a 5G core network, or an S1 interface for control plane signaling and user plane signaling communication when connected to an evolved packet core (EPC) network. The base stations 121 and 122 may communicate over an Xn interface using an Xn application protocol (XnAP) at 106, or communicate over an X2 interface using an X2 application protocol (X2AP) to exchange user plane and control plane data. The user equipment 110 may be connected to a public network (such as the Internet 160) via the core network 150 to interact with a remote service 170. Remote services 170 represent computing, communications, and storage devices used to provide any of a variety of services, including interactive voice or video communications, file transfers, streaming voice or video, and other technical services implemented in any manner, such as voice calls, video calls, website access, messaging services (e.g., text messaging or multimedia messaging), photo file transfers, enterprise software applications, social media applications, video games, streaming video services, and podcasts. Example Device

[0017] Figure 2 An example device diagram 200 illustrating user equipment 110 and one of base stations 120 . Figure 3 An example device diagram 300 illustrates a core network server 302. User equipment 110, base station 120, and / or core network server 302 may include for clarity Figure 2 or Figure 3 Additional functions and interfaces omitted in .

[0018] The user equipment 110 includes an antenna 202, a radio frequency front end 204 (RF front end 204), a wireless transceiver (e.g., an LTE transceiver 206 and / or a 5G NR transceiver 208) for communicating with the base station 120 in the RAN 140. The RF front end 204 of the user equipment 110 can couple or connect the LTE transceiver 206 and the 5G NR transceiver 208 to the antenna 202 to facilitate various types of wireless communications. The antenna 202 of the user equipment 110 may include an array of multiple antennas configured similarly or differently to each other. The antenna 202 and the RF front end 204 may be tuned and / or tunable to one or more frequency bands defined by the 3GPP LTE and 5G NR communication standards and implemented by the LTE transceiver 206 and the 5G NR transceiver 208. Additionally, antenna 202, RF front end 204, LTE transceiver 206, and / or 5G NR transceiver 208 may be configured to support beamforming for transmission and reception of communications with base station 120. By way of example and not limitation, antenna 202 and RF front end 204 may be implemented to operate in sub-gigahertz bands, sub-6 GHZ bands, and / or above 6 GHz bands as defined by 3GPP LTE and 5G NR communication standards.

[0019] The user device 110 also includes a processor 210 and a computer readable storage medium 212 (CRM 212). The processor 210 can be a single-core processor or a multi-core processor composed of various materials such as silicon, polysilicon, high-K dielectrics, copper, etc. The computer readable storage medium described herein excludes propagation signals. The CRM 212 can include any suitable memory or storage device, such as a random access memory (RAM), static RAM (SRAM), dynamic RAM (DRAM), non-volatile RAM (NVRAM), read-only memory (ROM), or flash memory that can be used to store device data 214 of the user device 110. The device data 214 includes user data, multimedia data, beamforming codebooks, applications, neural network tables, and / or operating systems of the user device 110, which can be executed by the processor 212 to implement user plane communications, control plane signaling, and user interaction with the user device 110.

[0020] In some implementations, the computer-readable storage medium 212 includes a neural network table 216 that stores various architectures and / or parameter configurations for forming a neural network, such as, for example, but not limited to, parameters specifying a fully connected layer neural network architecture, a convolutional layer neural network architecture, a recursive neural network layer, a plurality of connected hidden neural network layers, an input layer architecture, an output layer architecture, a plurality of nodes utilized by the neural network, coefficients utilized by the neural network (e.g., weights and biases), kernel parameters, a plurality of filters utilized by the neural network, a span / pooling configuration used by the neural network, an activation function for each neural network layer, interconnections between neural network layers, neural network layers to be skipped, and the like. Thus, the neural network table 216 includes any combination of NN formation configuration elements (e.g., architectures and / or parameter configurations) that can be used to create a NN formation configuration (e.g., a combination of one or more NN formation configuration elements) that defines and / or forms a DNN. In some implementations, a single index value of the neural network table 216 maps to a single NN formation configuration element (e.g., a 1:1 correspondence). Alternatively or additionally, a single index value of the neural network table 216 maps to a NN formation configuration (e.g., a combination of NN formation configuration elements). In some implementations, the neural network table includes input characteristics for each NN formation configuration element and / or NN formation configuration, wherein the input characteristics describe properties of training data used to generate the NN formation configuration element and / or NN formation configuration as further described.

[0021] In some implementations, the CRM 212 may also include a user equipment neural network manager 218 (UE neural network manager 218). Alternatively or additionally, the UE neural network manager 218 may be implemented in whole or in part as hardware logic or circuitry that is integrated or separate from other components of the user equipment 110. The UE neural network manager 218 accesses the neural network table 216, such as by an index value, and forms a DNN using the NN formation configuration elements specified by the NN formation configuration. In an implementation, the UE neural network manager forms multiple DNNs to process wireless communications (e.g., downlink communications and / or uplink communications exchanged with the base station 120).

[0022] Figure 2The device diagram of the base station 120 shown in the figure includes a single network node (e.g., gNode B). The functions of the base station 120 may be distributed across multiple network nodes or devices and may be distributed in any manner suitable for performing the functions described herein. The base station 120 includes an antenna 252, a radio frequency front end 254 (RF front end 254), and one or more wireless transceivers (e.g., one or more LTE transceivers 256, and / or one or more 5GNR transceivers 258) for communicating with the terminal 110. The RF front end 254 of the base station 120 may couple or connect the LTE transceiver 256 and the 5GNR transceiver 258 to the antenna 252 to facilitate various types of wireless communications. The antenna 252 of the base station 120 may include an array of multiple antennas configured similarly or differently to each other. The antenna 252 and the RF front end 254 may be tuned and / or tunable to one or more frequency bands defined by the 3GPP LTE and 5GNR communication standards and implemented by the LTE transceiver 256 and / or the 5GNR transceiver 258. Additionally, the antenna 252 , the RF front end 254 , the LTE transceiver 256 , and / or the 5G NR transceiver 258 may be configured to support beamforming (such as massive-MIMO) for transmission and reception of communications with the UE 110 .

[0023] The base station 120 also includes a processor 260 and a computer readable storage medium 262 (CRM 262). The processor 260 may be a single core processor or a multi-core processor composed of various materials such as silicon, polysilicon, high-K dielectrics, copper, etc. The CRM 262 may include any suitable memory or storage device, such as a random access memory (RAM), static RAM (SRAM), dynamic RAM (DRAM), non-volatile RAM (NVRAM), read-only memory (ROM), or flash memory that can be used to store device data 264 of the base station 120. The device data 264 includes network scheduling data, radio resource management data, beamforming codebooks, applications, and / or operating systems of the base station 120, which can be executed by the processor 260 to enable communication with the user equipment 110.

[0024] The CRM 262 also includes a base station manager 266. Alternatively or additionally, the base station manager 266 can be implemented in whole or in part as hardware logic or circuitry that is integrated or separate from other components of the base station 120. In at least some aspects, the base station manager 266 configures the LTE transceiver 256 and the 5G NR transceiver 258 for communication with the user equipment 110, as well as for communication with a core network such as the core network 150.

[0025] The CRM 262 also includes a base station neural network manager 268 (BS neural network manager 268). Alternatively or additionally, the BS neural network manager 268 can be implemented in whole or in part as hardware logic or circuitry that is integrated or separate from other components of the base station 120. In at least some aspects, the BS neural network manager 268 selects a NN formation configuration (such as by selecting a combination of NN formation configuration elements) to be utilized by the base station 120 and / or the UE 110 to configure a deep neural network for processing wireless communications. In some implementations, the BS neural network manager receives feedback from the UE 110 and selects a neural network formation configuration based on the feedback. Alternatively or additionally, the BS neural network manager 268 receives neural network formation configuration directions from a core network 150 element via a core network interface 276 or an inter-base station interface 274 and forwards the neural network formation configuration directions to the UE 110.

[0026] The CRM 262 includes a training module 270 and a neural network table 272. In an implementation, the base station 120 manages the NN information configuration and deploys it to the UE 110. Alternatively or additionally, the base station 120 maintains the neural network table 272. The training module 270 teaches and / or trains the DNN using known input data. For example, the training module 270 trains the DNN for different purposes, such as processing communications transmitted on a wireless communication system (e.g., encoding downlink communications, modulating downlink communications, demodulating downlink communications, decoding downlink communications, encoding uplink communications, modulating uplink communications, demodulating uplink communications, decoding uplink communications). This includes training the DNN offline (e.g., when the DNN is not actively involved in processing communications) and / or online (e.g., when the DNN is actively involved in processing communications).

[0027] In an implementation, the training module 270 extracts learned parameter configurations from the DNN to identify NN formation configuration elements and / or NN formation configurations, and then adds and / or updates the NN formation configuration elements and / or NN formation configurations in the neural network table 272. The extracted parameter configurations include any combination of information that defines the behavior of the neural network, such as node connections, coefficients, active layers, weights, biases, pooling, etc.

[0028] The neural network table 272 stores a plurality of different NN formation configuration elements and / or NN formation configurations generated using the training module 270. In some implementations, the neural network table includes input characteristics for each NN formation configuration element and / or NN formation configuration, wherein the input characteristics describe properties of the training data used to generate the NN formation configuration element and / or NN formation configuration. For example, the input characteristics include, by way of example and not limitation, power information, signal to interference plus noise ratio (SINR) information, channel quality indicator (CQI) information, channel state information (CSI), Doppler feedback, frequency band, block error rate (BLER), quality of service (QoS), hybrid automatic repeat request (HARQ) information (e.g., first transmission error rate, second transmission error rate, maximum retransmissions), delay, radio link control (RLC), automatic repeat request (ARQ) metric, received signal strength (RSS), uplink SINR, timing measurement, error metric, UE capability, BS capability, power mode, Internet Protocol (IP) layer throughput, end-to-end delay, end-to-end packet loss rate, etc. Thus, the input features sometimes include layer 1, layer 2, and / or layer 3 metrics. In some implementations, a single index value of the neural network table 272 maps to a single NN formation configuration element (e.g., a 1:1 correspondence). Alternatively or additionally, a single index value of the neural network table 272 maps to a NN formation configuration (e.g., a combination of NN formation configuration elements).

[0029] In an implementation, the base station 120 synchronizes the neural network table 272 with the neural network table 216 so that the NN formation configuration elements and / or input characteristics stored in one neural network table are replicated in the second neural network table. Alternatively or additionally, the base station 120 synchronizes the neural network table 272 with the neural network table 216 so that the NN formation configuration elements and / or input characteristics stored in one neural network table represent complementary functionality in the second neural network table (e.g., NN formation configuration elements for transmitter path processing in the first neural network table, NN formation configuration elements for receiver path processing in the second neural network table).

[0030] The base stations 120 also include inter-base station interfaces 274 (such as Xn and / or X2 interfaces) that the base station manager 266 configures to exchange user planes, control planes, and other information between other base stations 120 to manage communications between the base stations 120 and the user equipment 110. The base stations 120 include core network interfaces 276 that the base station manager 266 configures to exchange user planes, control planes, and other information with core network functions and / or entities.

[0031] exist Figure 3In the embodiment of the present invention, the core network server 302 may provide all or part of the functions, entities, services and / or gateways in the core network 150. Each function, entity, service and / or gateway in the core network 150 may be provided as a service in the core network 150, distributed across multiple servers, or implemented on a dedicated server. For example, the core network server 302 may provide all or part of the services or functions such as user plane function (UPF), access and mobility management function (AMF), serving gateway (S-GW), packet data network gateway (P-GW), mobility management entity (MME), evolved packet data gateway (ePDG), etc. The core network server 302 is illustrated as being implemented on a single server including a processor 304 and a computer readable storage medium 306 (CRM 306). The processor 304 may be a single-core processor or a multi-core processor composed of various materials such as silicon, polysilicon, high-K dielectrics, copper, etc. CRM 306 may include any suitable memory or storage device, such as random access memory (RAM), static RAM (SRAM), dynamic RAM (DRAM), non-volatile RAM (NVRAM), read-only memory (ROM), a hard drive, or flash memory that may be used to store device data 308 for core network server 302. Device data 308 includes data used to support core network functions or entities and / or an operating system of core network server 302, which may be executed by processor 304.

[0032] The CRM 306 also includes one or more core network applications 310, which are implemented on the CRM 306 in one implementation (as shown). The one or more core network applications 310 can implement functions such as UPF, AMF, S-GW, P-GW, MME, ePDG, etc. Alternatively or additionally, the one or more core network applications 310 can be implemented in whole or in part as hardware logic or circuits integrated or separated from other components of the core network server 302.

[0033] The CRM 306 also includes a core network neural network manager 312 that manages NN formation configurations for processing communications exchanged between the UE 110 and the base station 120. In some implementations, the core network neural network manager 312 analyzes various parameters, such as current signal channel conditions (e.g., as reported by the base station 120, as reported by other wireless access points, as reported by the UE 110 (via the base station or other wireless access point)), capabilities at the base station 120 (e.g., antenna configuration, cell configuration, MIMO capabilities, wireless capabilities, processing capabilities), capabilities of the UE 110 (e.g., antenna configuration, MIMO capabilities, wireless capabilities, processing capabilities), and the like. For example, the base station 120 obtains the various parameters during communications with the UE and forwards these parameters to the core network neural network manager 312. The core network neural network manager selects a NN information configuration based on these parameters that improves the accuracy of the DNN processing communications. Improved accuracy means improved accuracy, such as lower bit errors, of outputs generated by the neural network relative to a neural network configured using another NN formation configuration. The core network neural network manager 312 then communicates the selected NN formation configuration to the base station 120 and / or the UE 110. In an implementation, the core network neural network manager 312 receives UE and / or BS feedback from the base station 120 and selects an updated NN formation configuration based on the feedback.

[0034] The CRM 306 includes a training module 314 and a neural network table 316. In an implementation, the core network server 302 manages the NN formation configuration and deploys it to multiple devices in the wireless communication system, such as UE 110 and base station 120. Alternatively or additionally, the core network server maintains the neural network table 316 outside the CRM 306. The training module 314 teaches and / or trains the DNN using known input data. For example, the training module 314 trains the DNN to process different types of pilot communications transmitted through the wireless communication system. This includes training the DNN offline and / or online. In an implementation, the training module 314 extracts learned NN formation configurations and / or learned NN formation configuration elements from the DNN and stores the learned NN formation configuration elements in the neural network table 316. Therefore, the NN formation configuration includes any combination of architectural configurations (e.g., node connections, layer connections) and / or parameter configurations (e.g., weights, biases, pooling) that define or influence the behavior of the DNN. In some implementations, a single index value of the neural network table 316 maps to a single NN formation configuration element (e.g., a 1:1 correspondence). Alternatively or additionally, a single index value of the neural network table 316 maps to a NN formation configuration (e.g., a combination of NN formation configuration elements).

[0035] In some implementations, the training module 314 of the core network neural network manager 312 generates NN formation configurations and / or NN formation configuration elements that are complementary to the NN formation configurations and / or NN formation configuration elements stored in the neural network table 216 at the UE 110 and / or the neural network table 272 at the base station 121. As an example, the training module 314 generates a neural network table 316 having NN formation configurations and / or NN formation configuration elements that have a high variation in architecture and / or parameter configuration relative to a medium and / or low variation used to generate the neural network table 272 and / or the neural network table 216. For example, the NN formation configurations and / or NN formation configuration elements generated by the training module 314 correspond to fully connected layers, full kernel size, frequent sampling and / or pooling, high weighted precision, etc. Therefore, the neural network table 316 sometimes includes a high precision neural network at the expense of increased processing complexity and / or time.

[0036] The NN formation configurations and / or NN formation configuration elements generated by the training module 270 sometimes have more fixed architectures and / or parameter configurations (e.g., fixed connection layers, fixed kernel size, etc.) and fewer variations relative to the NN formation configurations and / or NN formation configuration elements generated by the training module 314. The training module 270, for example, generates streamlined NN formation configurations (e.g., faster computation time, less data processing) relative to the NN formation configurations generated by the training module 314 to optimize or improve the performance of end-to-end network communications at the base station 121 and / or the UE 110. Alternatively or additionally, the NN formation configurations and / or NN formation configuration elements stored at the neural network table 216 at the UE 110 include more fixed architectures and / or parameter configurations relative to those NN formation configurations and / or NN formation configuration elements stored in the neural network table 316 and / or the neural network table 272, which reduce the requirements at the UE 110 relative to the base station 121 and / or the core network server 302 (e.g., computation speed, fewer data processing points, fewer computations, less power consumption, etc.). In an implementation, changes to the fixed (or flexible) architecture and / or parameter configuration at each neural network are based on the processing resources of the device targeted to form the corresponding DNN (e.g., processing power, memory constraints, quantization constraints (e.g., 8-bit versus 16-bit), fixed-point versus floating-point calculations, floating-point operations per second (FLOPS), power availability). Thus, a UE or access point with fewer processing resources relative to a core network server or base station receives a NN information configuration optimized for the available processing resources.

[0037] The neural network table 316 stores a plurality of different NN formation configuration elements generated using the training module 314. In some implementations, the neural network table includes input characteristics for each NN formation configuration element and / or NN formation configuration, wherein the input characteristics describe properties about the training data used to generate the NN formation configuration. For example, the input characteristics may include power information, SINR information, CQI, CSI, Doppler feedback, RSS, error metric, minimum end-to-end (E2E) delay, expected E2E delay, E2E QoS, E2E throughput, E2E packet loss rate, service cost, etc.

[0038] The CRM 302 also includes an end-to-end machine learning controller 318 (E2E ML controller 318). The E2E ML controller 318 determines an end-to-end machine learning configuration (E2E ML configuration) for processing information (such as QoS flows) exchanged via E2E communications. In an implementation, the E2E ML controller analyzes any combination of ML capabilities (e.g., supported ML architectures, number of layers supported, available processing power, memory limitations, available power budget, fixed-point processing versus floating-point processing, maximum kernel size capabilities, compute capabilities) of devices participating in the E2E communication. Alternatively or additionally, the E2E ML controller analyzes any combination of QoS requirements, QoS parameters, and / or QoS characteristics to determine an E2E ML configuration that satisfies the associated requirements, parameters, and / or characteristics. In some implementations, the E2E ML controller obtains metrics characterizing a current operating environment and analyzes the current operating environment to determine the E2E ML configuration. This includes determining an E2E ML configuration that includes an architecture configuration combined with a parameter configuration that defines the DNN, or determining an E2E ML configuration that includes only a parameter configuration for updating the DNN.

[0039] In determining an E2E ML configuration, the E2E ML controller sometimes determines a partitioning of the E2E ML configuration that distributes processing functions associated with the E2E ML configuration across multiple devices. For clarity, Figure 3 The end-to-end machine learning controller 318 is illustrated as being separate from the core network neural network manager 312, but in alternative or additional implementations, the core network neural network manager 312 includes functionality performed by the end-to-end machine learning controller 318, or vice versa. Figure 3 The core network server 302 is illustrated as implementing the E2E ML controller 318 , but alternative or additional devices may implement the E2E ML controller, such as the base station 120 and / or other network elements.

[0040] The core network server 302 also includes a core network interface 320 for communicating user plane, control plane, and other information with other functions or entities in the core network 150, the base station 120, or the UE 110. In an implementation, the core network server 302 communicates the NN information configuration to the base station 120 using the core network interface 320. The core network server 302 may alternatively or additionally receive feedback from the base station 120 and / or the UE 110 through the base station 120 using the core network interface 320.

[0041] Having described an example environment and example devices that can be used for neural network formation configuration feedback in wireless communications, consider now a discussion of configurable machine learning modules in accordance with one or more implementations. Configurable machine learning modules

[0042] Figure 4 An example machine learning module 400 is illustrated. The machine learning module 400 implements an adaptive set of algorithms that learn and identify patterns within data. The machine learning module 400 can be implemented using any combination of software, hardware, and / or firmware.

[0043] exist Figure 4 In the embodiment of the present invention, the machine learning module 400 includes a deep neural network 402 (DNN 402) having a group of connected nodes (e.g., neurons and / or perceptrons) organized into three or more layers. The nodes between the layers can be configured in various ways, such as a partially connected configuration in which a first subset of nodes in the first layer is connected to a second subset of nodes in the second layer, a fully connected configuration in which each node in the first layer is connected to each node in the second layer, and the like. Neurons process input data to produce continuous output values, such as any real number between 0 and 1. In some cases, the output value indicates how close the input data is to the expected category. Perceptrons perform linear classification, such as binary classification, on the input data. Nodes, whether neurons or perceptrons, can use various algorithms to generate output information based on adaptive learning. Using DNNs, the machine learning module 400 performs a variety of different types of analysis, including single linear regression, multiple linear regression, logistic regression, stepwise regression, binary classification, multi-class classification, multivariate adaptive regression splines, local estimation scatter plot smoothing, and the like.

[0044] In some implementations, the machine learning module 400 learns adaptively based on supervised learning. In supervised learning, the machine learning module receives various types of input data as training data. The machine learning module processes the training data to learn how to map the input to the desired output. As an example, the machine learning module 400 receives digital samples of a signal as input data, and learns how to map the signal samples to binary data reflecting the information embedded in the signal. As another example, the machine learning module 400 receives binary data as input data, and learns how to map the binary data to digital samples of the signal, wherein the binary data is embedded in the signal. During the training process, the machine learning module 400 uses labeled or known data as input to the DNN. The DNN uses nodes to analyze the input and generate corresponding outputs. The machine learning module 400 compares the corresponding output with the true value data, and adapts the algorithm implemented by the node to improve the accuracy of the output data. Then, the DNN applies the adapted algorithm to the unlabeled input data to generate the corresponding output data.

[0045] The machine learning module 400 uses statistical analysis and / or adaptive learning to map inputs to outputs. For example, the machine learning module 400 uses characteristics learned from training data to correlate unknown inputs with outputs that are statistically likely to be within a threshold range or value. This allows the machine learning module 400 to receive complex inputs and identify corresponding outputs. Some implementations train the machine learning module 400 on characteristics of communications transmitted through a wireless communication system (e.g., time / frequency interleaving, time / frequency deinterleaving, convolutional encoding, convolutional decoding, power levels, channel equalization, inter-symbol interference, orthogonal amplitude modulation / demodulation, frequency division multiplexing / demultiplexing, transmission channel characteristics). This allows the trained machine learning module 400 to receive samples of signals as input, such as samples of downlink signals received at a user equipment, and to recover information from the downlink signal, such as binary data embedded in the downlink signal.

[0046] exist Figure 4 In the embodiment, the DNN includes an input layer 404, an output layer 406, and one or more hidden layers 408 located between the input layer and the output layer. Each layer has an arbitrary number of nodes, wherein the number of nodes between layers can be the same or different. In other words, the input layer 404 can have the same number and / or a different number of nodes as the output layer 406, the output layer 406 can have the same number and / or a different number of nodes as the hidden layer 408, and so on.

[0047] Node 410 corresponds to one of several nodes included in input layer 404, wherein the nodes perform independent calculations. As further described, the node receives input data and processes the input data using an algorithm to generate output data. Sometimes, the algorithm includes weights and / or coefficients that change based on adaptive learning. Therefore, the weights and / or coefficients reflect the information learned by the neural network. In some cases, each node can determine whether to pass the processed input data to the next node. For illustration, after processing the input data, node 410 can determine whether to pass the processed input data to node 412 and / or node 414 of hidden layer 408. Alternatively or additionally, node 410 passes the processed input data to the node based on the layer connection architecture. This process can be repeated in multiple layers until the DNN generates output using the nodes of output layer 406.

[0048] Neural networks may also employ various architectures that determine what nodes within the neural network are connected, how data is advanced and / or retained in the neural network, what weights and coefficients are used to process input data, how data is processed, and so on. These various factors collectively describe the NN formation configuration. To illustrate, a recursive neural network, such as a long short-term memory (LSTM) neural network, forms loops between node connections in order to retain information from previous portions of the input data sequence. The recursive neural network then uses the retained information for subsequent portions of the input data sequence. As another example, a feedforward neural network passes information to pass connections forward without forming loops that retain information. Although described in the context of node connections, it should be understood that the NN formation configuration may include various parameter configurations that affect how the neural network processes input data.

[0049] The NN formation configuration of a neural network can be characterized by various architectures and / or parameter configurations. To illustrate, consider an example of a DNN implementing a convolutional neural network. Typically, a convolutional neural network corresponds to a type of DNN in which each layer processes data using a convolution operation to filter the input data. Thus, a convolutional NN formation configuration is characterized by, by way of example and not limitation, pooling parameters, kernel parameters, weights, and / or layer parameters.

[0050] The pooling parameters correspond to parameters of a pooling layer that specifies a dimensionality reduction of input data within a convolutional neural network. For illustration, a pooling layer may combine node outputs at a first layer into node inputs at a second layer. Alternatively or additionally, the pooling parameters specify how and where a data layer of a processing neural network pools data. For example, a pooling parameter indicating "max pooling" configures the neural network to pool by selecting a maximum value from a data group generated by nodes of the first layer, and using the maximum value as an input to a single node of the second layer. A pooling parameter indicating "average pooling" configures the neural network to generate an average value from a data group generated by nodes of the first layer, and using the average value as an input to a single node of the second layer.

[0051] The kernel parameters indicate the filter size (e.g., width and height) to be used when processing the input data. Alternatively or additionally, the kernel parameters specify the type of kernel method to be used when filtering and processing the input data. For example, a support vector machine corresponds to a kernel method that uses regression analysis to identify and / or classify data. Other types of kernel methods include Gaussian processing, canonical correlation analysis, spectral clustering methods, etc. Thus, the kernel parameters may indicate the filter size and / or the type of kernel method to be applied in a neural network.

[0052] The weight parameters specify the weights and biases used by the algorithm within the node to classify the input data. In an implementation, the weights and biases are learned parameter configurations, such as those generated from training data.

[0053] The layer parameters specify layer connections and / or layer types, such as a fully connected layer type indicating that each node in a first layer (e.g., output layer 406) is connected to each node in a second layer (e.g., hidden layer 408), a partially connected layer type indicating which nodes in the first layer are to be disconnected from the second layer, an activation layer type indicating which filters and / or layers within a neural network are to be activated, etc. Alternatively or additionally, the layer parameters specify a type of node layer, such as a normalization layer type, a convolutional layer type, a pooling layer type, etc.

[0054] Although described in the context of pooling parameters, kernel parameters, weight parameters, and layer parameters, it should be understood that other parameter configurations may be used to form a DNN without departing from the scope of the claimed subject matter. Thus, a NN formation configuration may include any other type of parameters that may be applied to a DNN that affect how the DNN processes input data to generate output data.

[0055] Some implementations configure the machine learning module 400 based on the current operating environment. To illustrate, consider a machine learning module that is trained to generate binary data from digital samples of a signal. The transmission environment often modifies the characteristics of a signal traveling through the environment. The transmission environment often changes, which affects how the environment modifies the signal. For example, a first transmission environment modifies the signal in a first manner, while a second transmission environment modifies the signal in a manner different from the first manner. These differences affect the accuracy of the output results generated by the machine learning module. For example, a neural network configured to process communications transmitted on a first transmission environment may generate errors (e.g., bit errors exceeding a threshold) when processing communications transmitted on a second transmission environment.

[0056] Various implementations generate and store NN formation configurations and / or NN formation configuration elements (e.g., various architectures and / or parameter configurations) for different transmission environments. For example, the base station 120 and / or the core network server 302 trains the machine learning module 400 using any combination of the BS neural network manager 268, the training module 270, the core network neural network manager 312, and / or the training module 314. Training can occur offline when no active communication exchange occurs, or online during active communication exchange. For example, the base station 120 and / or the core network server 302 can mathematically generate training data, access files storing training data, obtain real-world communication data, etc. The base station 120 and / or the core network server 302 then extracts various learned NN formation configurations and stores them in a neural network table. Some implementations store input characteristics with each NN formation configuration, where the input characteristics describe various attributes of the transmission environment corresponding to the corresponding NN formation configuration. In an implementation, the neural network manager selects a NN formation configuration and / or NN formation configuration element by matching the current transmission environment and / or the current operating environment with the input characteristics.

[0057] Having described a configurable machine learning module, consider now a discussion of deep neural networks in wireless communication systems in accordance with one or more implementations. Deep Neural Networks in Wireless Communication Systems

[0058] Wireless communication systems include various complex components and / or functions, such as Figure 1 Example environment 100, Figure 2 An example device diagram 200 and Figure 3 The example device diagram 300 depicts various devices and modules, and in some implementations, devices participating in a wireless communication system link a series of functions together to enable information to be exchanged over a wireless connection.

[0059] For demonstration purposes, now consider Figure 5, which illustrates an example block diagram 500 and an example block diagram 502, each of which depicts an example processing chain utilized by a device in a wireless communication system. For simplicity, the block diagrams illustrate high-level functionality, and it should be understood that the block diagrams can include information that has been removed for clarity. Figure 5 Additional features omitted in .

[0060] exist Figure 5 , block diagram 500 includes a transmitter block 504 and a receiver block 506. Transmitter block 504 includes a transmitter processing chain that proceeds from top to bottom. The transmitter processing chain starts with input data, which proceeds to a coding stage, followed by a modulation stage, and then a radio frequency (RF) analog transmission (Tx) stage. The coding stage can include any type and number of coding stages used by a device to transmit data over a wireless communication system.

[0061] To illustrate, the example encoding stage receives binary data as input and processes the binary data using various encoding algorithms to append information to the binary data, such as frame information. Alternatively or additionally, the encoding stage transforms the binary data, such as by applying forward error correction that adds redundancy to aid in information recovery at the receiver. As another example, the encoding stage converts the binary data into symbols.

[0062] The example modulation stage receives as input the output generated by the encoding stage and embeds the input onto the signal. For example, the modulation stage generates digital samples of the signal embedded with the input from the encoding stage. Thus, in the transmitter block 504, the encoding stage and the modulation stage represent a high-level transmitter processing chain, which sometimes includes low-level complex functions such as convolutional encoding, serial-to-parallel conversion, cyclic prefix insertion, channel coding, time / frequency interleaving, etc. The RF analog Tx stage receives the output from the modulation stage, generates an analog RF signal based on the modulation stage output, and transmits the analog RF signal to the receiver block 506.

[0063] Receiver block 506 uses a receiver processing chain to perform complementary processing relative to transceiver block 504. The receiver processing chain illustrated in receiver block 506 proceeds from top to bottom and includes an RF analog receive (Rx) stage followed by a demodulation stage and a decoding stage.

[0064] The RF analog Rx stage receives the signal transmitted by the transmitter block 504 and generates an input used by the demodulation stage. As an example, the RF analog Rx stage includes a down-conversion component and / or an analog-to-digital converter (ADC) to generate samples of the received signal. The demodulation stage processes the input from the RF analog Rx stage to extract data embedded on the signal (e.g., data embedded by the modulation stage of the transmitter block 504). For example, the demodulation stage recovers symbol and / or binary data.

[0065] The decoding stage receives inputs such as recovered symbols and / or binary data from the demodulation stage and processes the inputs to recover the transmitted information. For illustration, the decoding stage corrects data errors based on forward error correction applied at the transmitter block, extracts payload data from frames and / or time slots, etc. Thus, the decoding stage generates recovered information.

[0066] As described above, the transmitter and receiver processing chains illustrated by the transmitter block 504 and the receiver block 506 are simplified for clarity and may include multiple complex modules. Sometimes, these complex modules are dedicated to specific functions and / or conditions. For example, consider a receiver processing chain that processes orthogonal frequency division modulation (OFDM) transmissions. In order to recover information from OFDM transmissions, the receiver block sometimes includes multiple processing blocks, each of which is dedicated to a specific function, such as an equalization block that corrects distortion in the received signal, a channel estimation block that estimates transmission channel properties to identify the effects of scattering, power attenuation, etc. on the transmission, etc. At high frequencies, such as 5G mmW signals in the 6 GHz band, these blocks may be computationally expensive and / or expensive (e.g., requiring considerable processing power, requiring expensive hardware). In addition, implementing a block that generates an output with accuracy within a desired threshold sometimes requires more specific and less flexible components. To illustrate, an equalization block for a 5G mmW signal in the 6 GHz band may not be able to be performed with the same accuracy in other bands, so different equalization blocks for different bands are required, and the complexity of the corresponding device is increased.

[0067] Some implementations include DNNs in the transmit and / or receiver processing chains. In block diagram 502, transmitter block 508 includes one or more deep neural networks 510 (DNN 510) in the transmitter processing chain, while receiver block 512 includes one or more deep neural networks 514 (DNN 514) in the receiver processing chain.

[0068] For simplicity, the DNN 510 in the transmitter block 508 corresponds to the encoding and modulation levels of the transmitter block 504. However, it should be understood that the DNN can perform any high-level and / or low-level operations found within the transmitter processing chain. For example, a first DNN performs low-level transmitter-side forward error correction, a second DNN performs low-level transmitter-side convolutional encoding, etc. Alternatively or additionally, the DNN 510 performs high-level processing, such as end-to-end processing corresponding to the encoding and modulation levels of the transmitter block 508.

[0069] In a similar manner, the DNN 514 in the receiver block 512 performs receiver processing chain functionality (e.g., demodulation stage, decoding stage). The DNN 514 can perform any high-level and / or low-level operations found within the receiver processing chain, such as low-level receiver-side error correction, low-level receiver-side symbol recovery, high-level end-to-end demodulation and decoding, etc. Therefore, the DNN in the wireless communication system can be configured to replace high-level operations and / or low-level operations in the transmitter and receiver processing chains. Sometimes, the DNN performing high-level operations and / or low-level operations can be configured and / or reconfigured based on the current operating environment, as further described. This provides more flexibility and adaptability to the processing chain relative to more specific and less flexible components.

[0070] Some implementations use multiple DNNs to process communication exchanges on a wireless communication system, where each DNN has a corresponding purpose (e.g., uplink processing, downlink processing, uplink encoding processing, downlink decoding processing, etc.). For demonstration purposes, now consider Figure 6 , which illustrates an example operating environment 600 including UE 110 and base station 121. Figure 6 A base station 121 is shown, but a base station 121 may be used. Figure 2 An alternative base station is shown. In an implementation, UE 110 and base station 121 exchange communications with each other over a wireless communication system by processing the communications using multiple DNNs.

[0071] exist Figure 6 In the example, the base station neural network manager 268 of the base station 121 includes a downlink processing module 602 for processing downlink communications, such as for generating downlink communications for transmission to the UE 110. For illustration, the base station neural network manager 268 forms a deep neural network 604 (DNN 604) in the downlink processing module 602 using the NN formation configuration, as further described. In some examples, the DNN 604 corresponds to Figure 5 In other words, DNN 604 performs some or all of the transmitter processing functions for generating downlink communications.

[0072] Similarly, the UE neural network manager 218 of the UE 110 includes a downlink processing module 606, wherein the downlink processing module 606 includes a deep neural network 608 (DNN 608) for processing (received) downlink communications. In various implementations, the UE neural network manager 218 forms the DNN 608 using the NN formation configuration. Figure 6 In the example, DNN 608 corresponds to Figure 5514 of UE 110, wherein the deep neural network 606 of UE 110 performs some or all receiver processing functions for (received) downlink communications. Therefore, DNN 604 and DNN 608 perform processing that complements each other (e.g., encoding / decoding, modulation / demodulation).

[0073] DNN 604 and / or DNN 608 may include multiple deep neural networks, wherein each DNN is dedicated to a corresponding channel, a corresponding purpose, etc. As an example, base station 121 uses a first DNN in DNN 604 to process downlink control channel information, uses a second DNN in DNN 604 to process downlink data channel information, etc. As another example, UE 110 uses a first DNN in DNN 608 to process downlink control channel information, uses a second DNN in DNN 608 to process downlink data channel information, etc.

[0074] Base station 121 and / or UE 110 also uses a DNN to process uplink communications. In environment 600, UE neural network manager 218 includes an uplink processing module 610, wherein uplink processing module 610 includes a deep neural network 612 (DNN 612) for generating and / or processing uplink communications (e.g., encoding, modulation). In other words, uplink processing module 610 processes pre-transmission communications as part of processing uplink communications. For example, UE neural network manager 218 uses a NN formation configuration to form DNN 612. Sometimes, DNN 612 corresponds to Figure 5 510. Thus, DNN 612 performs some or all of the transmitter processing functions for generating uplink communications transmitted from UE 110 to base station 121.

[0075] Similarly, the uplink processing module 614 of the base station 121 includes a deep neural network 616 (DNN 616) for processing (received) uplink communications, wherein the base station neural network manager 268 forms the DNN 616 using the NN formation configuration, as further described. In the example, the DNN 616 of the base station 121 corresponds to Figure 5 514 and performs some or all receiver processing functionality for (received) uplink communications (such as uplink communications received from UE 110). Sometimes, DNN 612 and DNN 616 perform complementary functions to each other. Alternatively or additionally, uplink processing module 610 and / or uplink processing module 614 include multiple DNNs, each of which has a dedicated purpose (e.g., processing a corresponding channel, performing a corresponding uplink function, etc.). Figure 6DNNs 604, 608, 612, and 616 are illustrated as residing within respective neural network managers to represent that the neural network managers form the DNNs, and it should be understood that the DNNs may be formed external to the neural network managers (e.g., UE neural network manager 218 and base station neural network manager 268) within different components, processing chains, modules, etc.

[0076] Having described deep neural networks in wireless communication systems, consider now a discussion of signaling and control transactions on a wireless communication system that can be used to configure a deep neural network for downlink and uplink communications in accordance with one or more implementations. Signaling and control transactions for configuring deep neural networks

[0077] Figure 7 , 8 9 illustrates an example signaling and control transaction diagram between a base station, a user equipment, and / or a core network server according to one or more aspects of neural network formation configuration feedback in wireless communication with respect to a deep neural network. The signaling and control transactions may be performed by Figure 1 The base station 120 and the UE 110 or Figure 3 The core web server 302 uses Figure 1-6 For example, in some implementations, the core network server 302 performs various signaling and control actions performed by the base station 120, such as Figure 7 and 8 shown.

[0078] Figure 7 Figure 700 illustrates a first example of signaling and control transactions for neural network formation configuration feedback in wireless communication. As shown, at 705, UE 110 optionally indicates UE capabilities (e.g., capabilities supported by the UE). In some implementations, the UE capabilities include ML-related capabilities, such as maximum kernel size capability, memory limitations, computing capabilities, supported ML architectures, number of layers supported, available processing power, memory limitations, available power budget, and fixed-point processing versus floating-point processing.

[0079] At 710, the base station 121 determines a neural network formation configuration. In determining the neural network formation configuration, the base station analyzes any combination of information such as the channel type (e.g., downlink, uplink, data, control, etc.) processed by the deep neural network, transmission medium properties (e.g., power measurements, signal to interference plus noise ratio (SINR) measurements, channel quality indicator (CQI) measurements), coding schemes, UE capabilities, BS capabilities, etc. In some implementations, the base station 121 determines the neural network formation configuration based on the UE capabilities indicated at 705. Alternatively or additionally, the base station 121 obtains the UE capabilities from a networked storage device such as a server.

[0080] For example, base station 121 receives a message from UE 110 (not shown) indicating one or more capabilities of the UE, such as, by way of example and not limitation, connectivity information, dual connectivity information, carrier aggregation capability, downlink physical parameter values, uplink physical parameter values, supported downlink / uplink categories, inter-frequency handover, and ML capabilities (e.g., maximum kernel size capability, memory limit, computational capability, supported ML architectures, number of layers supported, available processing power, memory limit, available power budget, fixed-point processing versus floating-point processing). Base station 121 identifies UE capabilities that affect how the UE processes communications, and / or how the base station processes communications from the UE, based on the message, and selects a neural network formation configuration that has improved output accuracy relative to other neural network formation configurations.

[0081] In some implementations, the base station 121 selects the neural network formation configuration from a plurality of neural network formation configurations. Alternatively or additionally, the base station 121 selects the neural network formation configuration by selecting a subset of neural network architecture formation elements in a neural network table. Sometimes, the base station 121 analyzes a plurality of neural network formation configurations and / or a plurality of neural network formation configuration elements included in the neural network table and determines the neural network formation configuration by selecting and / or creating a neural network formation configuration that aligns with current channel conditions (such as by matching channel type, transmission medium properties, etc. with input characteristics), as further described. Alternatively or additionally, the base station 121 selects the neural network formation configuration based on network parameters, such as scheduling parameters (e.g., scheduling multi-user, multiple-input, multiple-output (MU-MIMO) for downlink communications, scheduling MU-MIMO for uplink communications).

[0082] At 715, the base station 121 communicates the neural network formation configuration to the UE 110. In some implementations, the base station transmits a message specifying the neural network formation configuration, such as by transmitting a message including a mapping to a configuration such as Figure 2The base station 121 transmits a message including an index value of an entry in the neural network table of the neural network table 216 of the base station 121. Alternatively or additionally, the base station transmits a message including a neural network parameter configuration (e.g., weight values, coefficient values, number of filters). In some cases, the base station 121 specifies a purpose and / or a processing assignment in the message, wherein the processing assignment indicates what channel and / or where in the processing chain the configured neural network is applied, such as downlink control channel processing, uplink data channel processing, downlink decoding processing, uplink encoding processing, etc. Thus, the base station can communicate the processing assignment using a neural network formation configuration.

[0083] In some implementations, the base station 121 communicates multiple neural network formation configurations to the UE 110. For example, the base station transmits a first message directing the UE to use a first neural network formation configuration for uplink encoding and a second message directing the UE to use a second neural network formation configuration for downlink decoding. In some scenarios, the base station 121 communicates multiple neural network formation configurations and corresponding processing assignments in a single message. As another example, the base station communicates multiple neural network formation configurations using different radio access technologies (RATs). For example, the base station may transmit a first neural network formation configuration for downlink communication processing to the UE 110 using a first RAT and / or carrier, and transmit a second neural network formation configuration for uplink communication processing to the UE 110 using a second RAT and / or carrier.

[0084] At 720, UE 110 forms a first neural network based on the neural network formation configuration. For example, UE 110 uses the index value communicated by the base station to access a neural network table to obtain the neural network formation configuration and / or neural network formation configuration elements. Alternatively or additionally, UE 110 extracts the neural network architecture and / or parameter configuration from the message. UE 110 then forms a neural network using the neural network formation configuration, the extracted architecture and / or parameter configuration, etc. In some implementations, the UE uses the first neural network to process all communications, while in other implementations, the UE uses the first neural network to process select communications based on the processing assignment.

[0085] At 725, the base station 121 communicates information based on the neural network formation configuration. Figure 6 , base station 121 uses a second neural network configured with functions complementary to the first neural network to process downlink communications. In other words, the second neural network uses a second neural network formation configuration complementary to the neural network formation configuration. Then, at 730, UE 110 uses the first neural network to recover information.

[0086] Figure 8The signaling and control transaction diagram 800 illustrates a second example of signaling and control transactions for neural network formation configuration feedback in wireless communication regarding a deep neural network. In some implementations, the signaling and control transaction diagram 800 represents a Figure 7 Continuation of signaling and control transaction diagram 700 .

[0087] As shown, at 805, base station 121 communicates with UE 110 based on the first neural network formation configuration. Similarly, at 810, UE 110 communicates with base station 121 based on the first neural network formation configuration. For example, base station 121 communicates with UE 110 by using Figure 6 The DNN 604 processes one or more downlink communications to communicate with the UE 110, and the UE 110 communicates with the UE 110 by using Figure 6 DNN 608 processes downlink communications received from base station 121 to communicate with base station 121.

[0088] In an implementation, DNN 604 and DNN 608 are formed based on a first neural network formation configuration as described. To illustrate, DNN 604 and DNN 608 perform complementary functions to each other, where the first neural network formation configuration specifies complementary functions of each deep neural network (e.g., the base station forms a first neural network using the first neural network formation configuration, and the UE forms a second neural network complementary to the first neural network by using a neural network formation configuration that is complementary to the first neural network formation configuration). The complementary functions performed by the deep neural networks allow each side of the exchanged communications to remain synchronized (e.g., accurately recover information). Thus, the first neural network formation configuration specifies any combination of base station-side neural network formation configurations, complementary user device-side neural network formation configurations, and / or general neural network formation configurations used by each device participating in the communication exchange.

[0089] As another example, UE 110 uses Figure 6 The DNN 612 processes one or more uplink communications to the base station 121, and the base station 121 uses Figure 6 1 and 110 to process uplink communications received from UE 110 using a deep neural network 614. Similar to downlink communications, some implementations form DNN 612 and deep neural network 614 based on a first neural network forming configuration (e.g., the UE forms a first neural network using a first neural network forming configuration, and the base station forms a second neural network having a neural network forming configuration that is complementary to the first neural network forming configuration). Thus, base station 121 and UE 110 communicate with each other based on the first neural network forming configuration by forming a deep neural network based on the first neural network forming configuration and processing communications with the deep neural network.

[0090] At 815, the base station generates base station metrics, such as metrics based on uplink communications received from UE 110. Similarly, at 820, UE 110 generates UE metrics, such as metrics based on downlink communications received from base station 121. Base station 121 and / or UE 110 may generate any type of metric, such as power measurements (e.g., RSS), error metrics, timing metrics, QoS, delay, etc.

[0091] At 825, UE 110 communicates the metric to base station 121. In an implementation, UE 110 processes the metric communication using a deep neural network based on a first neural network formation configuration. Alternatively or additionally, UE 110 processes the metric communication using a neural network formed using a second neural network formation configuration. Thus, as further described, in some implementations, UE 110 maintains a plurality of deep neural networks, wherein each deep neural network has a designated purpose and / or processing assignment (e.g., a first neural network for downlink control channel processing, a second neural network for downlink data channel processing, a third neural network for uplink control channel processing, a fourth neural network for uplink data channel processing). Sometimes, base station 121 communicates a plurality of neural network formation configurations for forming the plurality of deep neural networks to UE 110.

[0092] At 830, the base station 121 identifies a second neural network formation configuration based on the metrics at 830. In some implementations, the base station 121 identifies the second neural network formation configuration based on UE metrics, base station metrics, or any combination thereof. This includes identifying any combination of architectural changes and / or parameter changes to the neural network formation configuration as further described, such as a small change to the neural network formation configuration involving updating coefficient parameters to account for a change in a return metric (e.g., a SINR change, a Doppler feedback change, a power level change, a BLER change). Alternatively or additionally, identifying the second neural network formation configuration includes a large change, such as reconfiguring node and / or layer connections based on a metric such as a change in power state (e.g., a transition from a radio resource connected state to an idle state).

[0093] In some implementations, the base station 121 identifies the partial and / or incremental neural network formation configuration as a second neural network formation configuration, wherein the partial and / or incremental neural network formation configuration indicates a change to a full neural network formation configuration. For example, the full neural network formation configuration includes an architecture configuration and a parameter configuration for the neural network, and the partial and / or incremental neural network formation configuration specifies a change and / or update to the parameter configuration based on using the same architecture configuration indicated in the full neural network formation configuration.

[0094] In some implementations, the base station identifies the second neural network formation configuration by identifying a neural network formation configuration in a neural network table that improves the ability of the UE and / or the ability of the base station to recover data from the communication (e.g., improves the accuracy of the recovered information). Figure 2 The base station 121 may further include a base station neural network manager 268 to identify a neural network formation configuration that compensates for the problem identified by the UE measurements and / or the base station measurements. As another example, the base station 121 identifies a neural network formation configuration having one or more input characteristics that align with the changed operating conditions identified by the UE measurements and / or the base station measurements. Alternatively or additionally, the base station 121 identifies a neural network formation configuration that produces similar results but with less processing, such as for a scenario where the UE moves to a lower power state.

[0095] At 840, base station 121 directs UE 110 to update the first neural network using the second neural network formation configuration. For example, the base station generates a configuration including Figure 2 In some implementations, base station 121 indicates in the message an update message of an index value of a second neural network formation configuration in neural network table 216 of UE 110. In some implementations, base station 121 indicates in the message a time instance that directs UE 110 when to apply the second neural network formation configuration. In other words, the time instance directs UE 110 to switch from processing communications using the first neural network formation configuration to processing communications using the second neural network formation configuration at the time specified in the time instance. In an implementation, the base station transmits updates to the downlink neural network using a first carrier or RAT and transmits updates to the uplink neural network using a second carrier or RAT.

[0096] At 845, the base station 121 updates the base station neural network based on the second neural network forming configuration, such as a deep neural network formed based on the first neural network forming configuration and used to communicate with the UE 110 at 805 (e.g., a deep neural network for processing downlink communications, a deep neural network for processing uplink communications). Similarly, at 850, the UE 110 updates the user equipment neural network based on the second neural network forming configuration, such as a deep neural network that performs a complementary function to the base station neural network updated at 845. As an example, the UE extracts an index value and / or a time value from the update message transmitted by the base station at 840. The UE 110 obtains the second neural network forming configuration and modifies the user equipment neural network at the time specified in the update message. Thus, the UE 110 processes communications using the first neural network forming configuration until the specified time in the update message, at which time the UE 110 switches to processing communications using the second neural network forming configuration.

[0097] In an implementation, the base station 121 and / or the UE 110 iteratively performs Figure 8 The signaling and control transactions represented by dashed lines in FIG. 8 process the signaling and control transactions depicted in diagram 800. These iterations allow base station 121 and / or UE 110 to dynamically modify the communication processing chain based on changing operating conditions, as further described.

[0098] Figure 9-1 and Figure 9-2 Figure 900 illustrates a third example of signaling and control transactions for neural network formation configuration feedback in wireless communication. Figure 9-1 As shown, at 905, the core network server 302 determines a neural network formation configuration based on various metrics and / or parameters, such as metrics from the UE 110, metrics from the base station 121, UE capabilities, etc. For example, the core network server receives any combination of metrics and / or parameters from the base station 121 and / or the UE 110, such as power information, SINR information, CQI, CSI, Doppler feedback, QoS, delay, UE capabilities, base station type (e.g., eNB, gNB, or ng-eNB), protocol version, error metrics, UE capabilities, BS capabilities, power mode, etc. The core network server 302 then determines the neural network formation configuration based on the metrics, parameters, etc.

[0099] In some implementations, the core network server 302 determines a plurality of different neural network formation configurations, each neural network formation configuration being specific to a respective base station and / or respective UE. Alternatively or additionally, the core network server 302 determines a neural network formation configuration used by a plurality of base stations and / or UEs. Sometimes, the core network server 302 determines a default neural network formation configuration used by the base stations and / or UEs to initially connect to each other. As further described, the default neural network formation configuration corresponds to a general neural network formation configuration that configures a deep neural network to process various input data and / or channel conditions with an accuracy within a threshold range or value. However, a dedicated neural network formation configuration corresponds to a deep neural network that is tuned to a particular type of input data and / or a particular channel condition. In determining the neural network formation configuration, some implementations of the core network server determine a complementary neural network formation configuration (e.g., a base station side neural network formation configuration that is complementary to a user equipment side neural network formation configuration).

[0100] At 910, the core network server 302 communicates the neural network formation configuration to the base station 121. For example, the core network server 302 communicates the neural network formation configuration to the base station 121. Figure 3 The core network interface 320 communicates the index value to the base station 121, where the index value is mapped to Figure 2Entries in the neural network table 272. Alternatively or additionally, the core network server 302 communicates various parameter configurations, such as coefficients, weights, layer connection values, etc.

[0101] At 915, the base station 121 forwards the neural network formation configuration to the UE 110. For example, the base station wirelessly transmits the index value to the UE, wirelessly transmits the parameter configuration to the UE, etc. When communicating the neural network formation configuration, the core network server 302 and / or the base station 121 sometimes indicates the processing assignment (e.g., downlink data channel processing, uplink control channel processing, decoding processing, uplink encoding processing, uplink modulation processing, downlink demodulation processing) of the neural network formation configuration.

[0102] In some implementations, base station 121 adds modifications to the neural network formation configuration before forwarding to UE 110. For example, base station 121 may have access to more updated information than the core network server, such as through UE capabilities received from UE 110. Base station 121 sometimes adapts the neural network formation configuration based on updated information (e.g., UE capabilities specific to UE 110), such as by removing layers and / or nodes to reduce the corresponding complexity of the deep neural network at UE 110 based on available processing power, battery power, available radio, etc. at the UE. As another example, base station 121 adds a convolutional layer to the neural network formation configuration based on the updated information. Base station 121 then forwards the modified neural network formation configuration to UE 110 in place of the neural network formation configuration received from core network server 302.

[0103] At 920, base station 121 uses the neural network formation configuration to form a first deep neural network, such as by identifying the base station side neural network formation configuration and forming the first deep neural network using the base station side neural network formation configuration. For illustration, base station 121 accesses Figure 2 Similarly, at 925, UE 110 uses the neural network formation configuration to form a second deep neural network, such as by identifying a complementary and / or user device side neural network formation configuration, and using the complementary and / or user device side neural network formation configuration to form the second neural network. In some implementations, UE 110 accesses the second neural network by using an index value. Figure 2 The neural network table 216 is used to obtain a complementary and / or user device side neural network formation configuration. Therefore, the first deep neural network and the second deep neural network are neural networks synchronized based on the neural network formation configuration determined and communicated by the core network server at 905 and 910.

[0104] exist Figure 9-2, at 930, base station 121 communicates with UE 110 using a first deep neural network, such as by generating and / or processing downlink communications to UE 110, by receiving and / or processing uplink communications from UE 110, etc. Similarly, at 935, UE 110 communicates with base station 121 using a second deep neural network at 935. In other words, UE 110 communicates with base station 121 using a complementary deep neural network (e.g., a second deep neural network) based on a neural network formation configuration as further described.

[0105] At 940, the base station generates base station metrics, wherein the base station 121 generates the metrics based on the communication at 930 and / or 935. Thus, the base station metrics may be based on uplink communications received from the UE 110. For example, the base station 121 generates uplink received power, uplink SINR, uplink packet error, uplink throughput, timing measurements, etc. In some implementations, the base station includes base station capabilities (BS capabilities) in the base station metrics, such as processing power (e.g., macro base station, small cell base station), power status, etc.

[0106] Similarly, at 945 , UE 110 generates UE metrics (eg, power information, SINR information, CQI, CSI, Doppler feedback, QoS, delay) based on downlink communications from base station 121 , and communicates the UE metrics to base station 121 at 950 .

[0107] At 955, the base station 121 forwards the metrics to the core network server 302, such as via the core network interface 320. This includes any combination of the base station metrics and the UE metrics generated at 940 and / or 945. Then, at 960, the core network server 302 determines updates to the neural network formation configuration. This may include any combination of architectural structure changes (e.g., reconfiguring node connections, reconfiguring active / inactive layers), changing applied processing parameters (e.g., coefficients, kernels), etc. Thus, the core network server 302 sometimes identifies small changes and / or large changes, such as reference Figure 8 830 of the base station 121. By receiving the metrics generated from the UE 110 via the base station 121, the core network server 302 receives feedback about the communication channel between the base station and the UE and / or an indication of how well the neural network handles communications transmitted on the communication channel. The core network server 302 analyzes the feedback to identify adjustments to the neural network formation configuration (e.g., selecting a configuration that reduces the processing time of the deep neural network) that, when applied to the neural network, improves how the deep neural network processes communications (e.g., more accurate data recovery, less erroneous data) and / or how efficiently the neural network processes communications.

[0108] In an implementation, the core network server 302, the base station 121 and / or the UE 110 iteratively performs the signaling and control transactions described in the signaling and control transaction diagram 900, which is composed of Figure 9-1 and 9-2 denoted by a dashed line from 960 back to 910. These iterations allow the core network server 302, base station 121, and / or UE 110 to dynamically modify the communication processing chain based on changing operating conditions, as further described.

[0109] In some implementations, the core network server 302 receives feedback from multiple UEs and / or base stations in the wireless communication system. This provides the core network server with a larger view of how well the wireless communication system is performing, what devices are communicating through the wireless communication system, how well these devices are communicating, etc. In various implementations, the core network server 302 determines updates to the neural network formation configuration based on optimizing communication for multiple devices and / or overall system performance rather than optimizing communication for a specific device.

[0110] Having described signaling and control transactions that can be used to configure a neural network to handle communications, consider now some example methods in accordance with one or more implementations. Example Method

[0111] refer to Fig.10 and Fig.11 Example methods 1000 and 1100 are described for forming configuration feedback according to one or more aspects of a neural network in wireless communication. The order in which the method blocks are described is not intended to be construed as limiting, and any number of the described method blocks may be skipped or combined in any order to implement a method or an alternative method. In general, any components, modules, methods, and operations described herein may be implemented using software, firmware, hardware (e.g., fixed logic circuitry), manual processing, or any combination thereof. Some operations of the example methods may be described in the general context of executable instructions stored on a computer-readable storage memory local and / or remote to a computer processing system, and implementations may include software applications, programs, functions, and the like. Alternatively or additionally, any functions described herein may be performed at least in part by one or more hardware logic components, such as, but not limited to, a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), an application specific standard product (ASSP), a system on a chip (SoC), a complex programmable logic device (CPLD), and the like.

[0112] Fig.10An example method 1000 for configuring a neural network to process communications exchanged over a wireless communication system is illustrated. In some implementations, the operations of the method 1000 are performed by a network entity, such as any of the base station 120 or the core network server 302.

[0113] At 1005, a network entity determines a neural network formation configuration for a deep neural network for processing communications transmitted over a wireless communication system. For example, the network entity (e.g., base station 121, core network server 302) determines the neural network formation configuration based at least in part on metrics, feedback, and / or other types of information from a user device (e.g., UE 110). For illustration, base station 121 receives a message from UE 110 indicating one or more capabilities of UE 110. Base station 121 then determines the neural network formation configuration based at least in part on the capabilities received from UE 110. Alternatively or additionally, base station 121 forwards the UE capabilities to core network server 302, and core network server 302 determines the neural network formation configuration based on the capabilities of UE 110. As another example, base station 121 determines the neural network formation configuration based on scheduling MU-MIMO downlink and / or uplink transmissions in the wireless communication system. In determining the neural network formation configuration, the network selects any combination of neural network formation configuration elements for a user device-side deep neural network, a base station-side deep neural network, and / or a deep neural network corresponding to both a user device-side deep neural network and a base station-side deep neural network.

[0114] When determining the neural network formation configuration, the network entity (e.g., base station 121, core network server 302) sometimes selects a default neural network formation configuration. Alternatively or additionally, the network entity analyzes multiple neural network formation configurations and selects a neural network formation configuration from the multiple neural network formation configurations that is aligned with current channel conditions, capabilities of a particular UE, etc.

[0115] At 1010, a network entity generates a message including an indication of a neural network formation configuration for a deep neural network. For example, the network entity (e.g., base station 121, core network server 302) generates a message including an index value that maps to one or more entries of a neural network table (e.g., neural network table 216, neural network table 272, neural network table 316). In some implementations, the network entity includes an indication of a processing assignment for the deep neural network, wherein the processing assignment specifies a processing chain function for a deep neural network formed using the neural network formation configuration. Alternatively or additionally, the network entity specifies in the message a time instance indicating a time and / or location to begin processing communications with the deep neural network.

[0116] At 1015, the network entity transmits a message to the user equipment to direct the user equipment to form a deep neural network using the neural network formation configuration and to process communications transmitted over the wireless communication system using the deep neural network. For example, the network entity (e.g., core network server 302) transmits the message to the user equipment (e.g., UE 110) by communicating the message to base station 121 (which transmits the message to the user equipment). Alternatively or additionally, the network entity (e.g., base station 121) communicates the message to the user equipment (e.g., UE 110). In an implementation, the deep neural network corresponds to a user equipment side deep neural network.

[0117] In some implementations, a network entity (e.g., base station 121) transmits a message using a specific RAT and / or carrier based on the processing assignment of the deep neural network. For example, base station 121 transmits a first neural network formation configuration having a first processing assignment (e.g., downlink data channel processing) using a first RAT and / or carrier, and transmits a second neural network formation configuration having a second processing assignment (e.g., downlink control channel processing) using a second RAT and / or carrier.

[0118] In some implementations, at 1020, the network entity analyzes the feedback received from the user equipment. As an example, the network entity (e.g., base station 121) receives metrics from the user equipment (e.g., UE 110) and analyzes the metrics to determine whether to update and / or reconfigure the neural network with a different neural network formation configuration. Alternatively or additionally, the base station forwards the feedback to the network entity (e.g., core network server 302), and the network entity analyzes the feedback to determine whether to update and / or reconfigure the neural network with a different neural network formation configuration.

[0119] At 1025, the network entity updates the deep neural network based on the feedback. For example, the network entity (e.g., base station 121, core network server 302) analyzes a plurality of neural network formation configurations included in a neural network table (e.g., neural network table 216, neural network table 272, neural network table 316) and selects a second neural network formation configuration that is aligned with the new channel condition indicated by the feedback. The network entity generates a second message including an indication of the second neural network formation configuration and transmits the second message to the user equipment. This includes configuring and transmitting the second message as described at 1010 and / or 1015. In an implementation, the network entity iteratively performs the various operations of method 1000 indicated herein with a dashed line from 1025 back to 1010.

[0120] Fig.11An example method 1100 for forming a neural network based on a neural network formation configuration is illustrated. In some aspects, the operations of method 1100 are implemented by a UE, such as UE 110.

[0121] At 1105, a user equipment receives a message indicating a neural network formation configuration of a deep neural network for processing communications transmitted via a wireless communication system. For example, a user equipment (e.g., UE 110) receives a message from a base station (e.g., base station 121) indicating the neural network formation configuration and / or a processing assignment for a deep neural network formed using the neural network formation configuration. In some implementations, the user equipment receives multiple messages, each message relating to a different neural network formation configuration (e.g., a first message indicating a first neural network formation configuration for a first deep neural network for processing downlink communications, a second message indicating a second neural network formation configuration for a second deep neural network for processing uplink communications). In an implementation, the user equipment receives the message as a broadcast message from a base station or as a UE-specific message.

[0122] At 1110, the user device forms a deep neural network using the neural network formation configuration indicated in the message. To illustrate, the user device (e.g., UE 110) extracts parameter configurations (such as coefficients, weights, layer connections, etc.) from the message and forms the deep neural network using the extracted parameter configurations. As another example, the user device (e.g., UE 110) extracts an index value from the message and obtains the parameter configuration from a neural network table, as further described. Alternatively or additionally, the user device extracts a time instance from the message, wherein the time instance indicates a time and / or location to begin processing communications using a deep neural network formed using the neural network formation configuration.

[0123] At 1115, the user equipment receives a communication from the base station. For example, the user equipment (e.g., UE 110) receives a downlink data channel communication from the base station (e.g., base station 121), receives a downlink control channel communication from the base station, etc. At 1120, the user equipment processes the communication using a deep neural network to extract information transmitted in the communication. As an example, the user equipment (e.g., UE 110) processes the communication using a deep neural network based on a processing assignment included in the message, such as by demodulating and / or decoding downlink communications from a base station (e.g., base station 121), encoding and / or modulating uplink communications to the base station, etc. using a deep neural network.

[0124] In some implementations, the user equipment optionally transmits feedback based on the communication at 1125. For example, the user equipment (e.g., UE 110) generates metrics (e.g., error metrics, SINR information, CQI, CSI, Doppler feedback) based on the communication and transmits these metrics as feedback to the base station (e.g., base station 121).

[0125] Having described some example methods that can be used to implement various aspects of base station-user equipment messaging using deep neural networks, consider now a discussion of generating and communicating neural network formation configurations according to one or more implementations. Generate and communicate neural network configurations

[0126] In supervised learning, a machine learning module processes labeled training data to generate an output. The machine learning module receives feedback about the accuracy of the generated output and modifies processing parameters to improve the accuracy of the output. Fig.12 An example 1200 is illustrated describing various aspects of generating multiple NN formation configurations. Sometimes, various aspects of example 1200 are represented by Figure 2 and Figure 3 This can be implemented by any combination of the training module 270, the base station neural network manager 268, the core network neural network manager 312 and / or the training module 314.

[0127] Fig.12 The upper part includes Figure 4 400 of the machine learning module. In an implementation, the neural network manager determines to generate different NN formation configurations. To illustrate, consider a scenario in which the base station neural network manager 268 determines to generate a NN formation configuration by selecting a combination of NN formation configuration elements from a neural network table, wherein the NN formation configuration corresponds to a UE decoding and / or demodulating downlink communications. In other words, the NN formation configuration (through a combination of NN formation configuration elements) forms a DNN that processes downlink communications received by the UE. However, transmission channel conditions often change, which in turn affects the characteristics of downlink communications. For example, a first transmission channel distorts downlink communications by introducing a frequency offset, a second transmission channel distorts downlink communications by introducing a Doppler effect, a third transmission channel distorts downlink communications by introducing a multipath channel effect, and so on. In order to accurately process downlink communications (e.g., reduce bit errors), various implementations select multiple NN formation configurations, wherein each NN formation configuration (and associated combination of NN formation configuration elements) corresponds to a corresponding input condition, such as a first transmission channel, a second transmission channel, and the like.

[0128] Training data 1202 represents example inputs to machine learning module 400. Fig.12, the training data represents data corresponding to the downlink communication. For example, the training data 1202 can include digital samples of the downlink communication signal, recovered symbols, recovered frame data, etc. In some implementations, the training module mathematically generates the training data or accesses a file storing the training data. At other times, the training module obtains real-world communication data. Therefore, the training module can use mathematically generated data, static data, and / or real-world data to train the machine learning module. Some implementations generate input characteristics 1204 that describe various qualities of the training data, such as transmission channel metrics, UE capabilities, UE speed, etc.

[0129] The machine learning module 400 analyzes the training data and generates an output 1206, represented here as binary data. Some implementations use the same training data set and / or additional training data sets with the same input characteristics to iteratively train the machine learning module 400 to improve the accuracy of the machine learning module. During training, the machine learning module modifies some or all of the architecture and / or parameter configurations of the neural network included in the machine learning module, such as node connections, coefficients, kernel size, etc. At a certain point in the training, the training module determines to extract the architecture and / or parameter configuration 1208 (e.g., pooling parameters, kernel parameters, layer parameters, weights) of the neural network, such as when the training module determines that the accuracy meets or exceeds the expected threshold, the training process meets or exceeds the number of iterations, etc. The training module then extracts the architecture and / or parameter configuration from the machine learning module for use as a NN formation configuration and / or NN formation configuration element. The architecture and / or parameter configuration can include any combination of fixed architecture and / or parameter configuration, and / or variable architecture and / or parameter configuration.

[0130] Fig.12 The lower portion of includes a neural network table 1212 representing a set of NN formation configuration elements, such as Figure 2 and Figure 3 1202. The neural network table 1212 stores various combinations of architecture configurations, parameter configurations, and input characteristics, although alternative implementations exclude input characteristics from the table. Various implementations update and / or maintain NN formation configuration elements and / or input characteristics as the machine learning module learns additional information. For example, at index 1214, the neural network manager and / or training module updates the neural network table 1212 to include the architecture and / or parameter configuration 1208 generated by the machine learning module 400 when analyzing the training data 1202.

[0131] The neural network manager and / or training module may alternatively or additionally add the input characteristics 1204 to the neural network table and link the input characteristics to the architecture and / or parameter configuration 1208. This allows the input characteristics to be obtained at the same time as the architecture and / or parameter configuration is obtained, such as by using an index value that references the neural network table (e.g., references the NN formation configuration, references the NN formation configuration element). In some implementations, the neural network manager selects the NN formation configuration by matching the input characteristics with the current operating environment (such as by matching the input characteristics with current channel conditions, UE capabilities, UE characteristics (e.g., speed, location, etc.), etc.).

[0132] Having described generating and communicating a neural network formation configuration, consider now a discussion of signaling and control transactions over a wireless communication system that can be used to communicate a neural network formation configuration in accordance with one or more implementations. Signaling and control transactions used to communicate the configuration of neural network formation

[0133] Figures 13 to 15 An example signaling and control transaction diagram between a base station, a user equipment, and / or a core network server according to one or more aspects of communicating a neural network form configuration (such as communicating a NN form configuration) is illustrated. In an implementation, the signaling and control transactions may be performed by Figure 1 Any one of the base station 120 and the UE 110, or Figure 3 The core web server 302 uses Figures 1 to 12 components to execute.

[0134] A first example of a signaling and control transaction for communicating a neural network formation configuration is represented by Fig.13 In an implementation, some or all of the signaling and control transactions described in the reference signaling and control transaction diagram correspond to the reference signaling and control transaction diagram. Figure 7 and / or Figure 8 Signaling and control transactions described.

[0135] As shown, at 1305, the base station 121 maintains a neural network table. For example, the base station neural network manager and / or a training module of the base station 121 (e.g., the base station neural network manager 268, the training module 270) generates and / or maintains a neural network table (e.g., the neural network table 272) using any combination of mathematically generated training data, data extracted from real-world communications, files, etc. In various implementations, the base station 121 maintains a plurality of neural network tables, wherein each neural network table includes a plurality of neural network formation configurations and / or neural network formation configuration elements for a specified purpose, such as a first neural network table specified for data channel communications, a second neural network table specified for control channel communications, etc.

[0136] At 1310, base station 121 transmits a neural network table to UE 110. As an example, the base station transmits the neural network table using layer 3 messaging (e.g., a radio resource control (RRC) message). When transmitting the neural network table, the base station transmits any combination of an architecture and / or parameter configuration that can be used to form a deep neural network, examples of which are provided in the present disclosure. Alternatively or additionally, the base station transmits an indication of processing assignments specified for the neural network table along with the neural network table. Thus, the base station transmits multiple neural network tables to the UE, with corresponding processing assignments specified for each neural network table. In some implementations, base station 121 broadcasts the neural network table to a group of UEs. At other times, base station 121 transmits a UE-specific neural network table to UE 110.

[0137] At 1315, base station 121 identifies a neural network formation configuration to use when processing communications. For example, the base station selects a combination of neural network formation architecture elements (such as in Figure 7 710 of ), by analyzing any combination of information (such as the channel type (e.g., downlink, uplink, data, control, etc.), transmission medium properties (e.g., power measurements, signal to interference plus noise ratio (SINR) measurements, channel quality indicator (CQI) measurements), coding schemes, UE capabilities, BS capabilities, etc.) processed by the deep neural network to determine the neural network formation configuration to use when processing communications. Thus, in some implementations, base station 121 determines the neural network formation configuration to use when processing communications based on the various metrics and / or parameters received (such as in Figure 8 815, 820, and 830 of 10101 ) to identify the neural network formation configuration. In some implementations, base station 121 identifies a default neural network formation configuration to use as the neural network formation configuration. Thus, identifying the neural network formation configuration can include identifying the default neural network formation configuration and / or an update to the neural network formation configuration.

[0138] When identifying a neural network formation configuration, the base station 121 finds a neural network formation configuration in the neural network table that corresponds to the determined neural network formation configuration. In other words, the base station 121 identifies the neural network formation configuration that is aligned with the determined neural network formation configuration, such as by correlating and / or matching input characteristics. Figure 2 The base station may further include a neural network formation configuration and / or a neural network formation configuration element in the neural network table 272 and / or the neural network table 216. When identifying the neural network formation configuration and / or the neural network formation configuration element in the neural network table, the base station may identify the index value of the neural network formation configuration and / or the neural network formation configuration element.

[0139] At 1320, base station 121 transmits an instruction directing UE 110 to form a deep neural network using a neural network formation configuration from the neural network table. Figure 7 715 of and / or in Figure 8 840 of , the base station 121 communicates the index value to the UE 110 and directs the UE 110 to form a deep neural network using the neural network formation configuration indicated by the index value. The base station can transmit the indication to the UE in any suitable manner. As an example, the base station uses a layer 2 message (e.g., a radio link control (RLC) message, a media access control (MAC) control element) to transmit the index value corresponding to the neural network formation configuration. In some implementations, the base station compares the current operating environment (e.g., one or more of channel conditions, UE capabilities, BS capabilities, metrics) with the input characteristics stored in the neural network table and identifies the stored input characteristics that align with the current operating environment. Then, the base station obtains the index value of the stored input characteristic, which in turn provides the index value of the neural network formation configuration and / or the neural network formation configuration element. The base station 121 then transmits the index value as an indication. Sometimes, the base station 121 includes a processing assignment to indicate a location in the processing chain to apply the deep neural network. In some implementations, the base station uses a downlink control channel to transmit the index value and / or the processing assignment.

[0140] Sometimes, the base station transmits a rule to the UE that specifies an operating parameter related to applying a neural network formation configuration. In one example, the rule includes a time instance indicating when to process a communication with a deep neural network formed using the neural network formation configuration. Alternatively or additionally, when the data channel and the control channel are within the time threshold, the rule specifies a time threshold that directs the UE to use a default neural network formation configuration instead of a specified neural network formation configuration. Additionally or alternatively, when the data channel and the control channel are within the time threshold, the rule may direct the user equipment to use the same neural network formation configuration for data channel communication and control channel communication. To illustrate, consider an example in which the UE processes data channel communication using a first deep neural network (formed using the first neural network formation configuration) and processes control channel communication using a second deep neural network (formed using the second neural network formation configuration). If the data channel communication and the control channel communication fall within the time threshold specified by the time instance, the UE processes both channels using a default deep neural network (formed using the default neural network formation configuration) and / or the same deep neural network because there may not be enough time to switch between the first deep neural network and the second deep neural network.

[0141] Sometimes, the base station communicates a default neural network formation configuration using a downlink control channel to specify the default neural network formation configuration to the UE, wherein the default neural network formation configuration forms a deep neural network that processes various input data. In some implementations, the default neural network formation configuration forms a deep neural network that processes various input data with an accuracy within a threshold range. The default neural network formation configuration can include a general neural network formation configuration.

[0142] To illustrate, some implementations generate or select neural network formation configurations for specific operating conditions, such as a first neural network formation configuration specific to UE downlink control channel processing (e.g., demodulation and / or decision) utilizing a current operating environment “X”, a second neural network formation configuration specific to UE downlink control channel processing utilizing a current operating environment “Y”, etc. For example, the first neural network formation configuration can be associated with a current operating environment in which a detected interference level is high, the second neural network formation configuration can be associated with a current operating environment in which a detected interference level is low, the third neural network formation configuration can be associated with a current operating environment in which a connected UE appears to be stationary, the fourth neural network formation configuration can be associated with a current operating environment in which a connected UE appears to be moving and has a particular speed, etc.

[0143] Forming a deep neural network using a neural network formation configuration for a particular operating condition (relative to forming a deep neural network using other neural network formation configurations) improves the accuracy of outputs generated by the deep neural network when processing input data corresponding to the particular operating condition. However, this introduces a tradeoff because the deep neural network formed using the neural network formation configuration for the particular operating condition generates outputs with lower accuracy when processing inputs associated with other operating conditions. In contrast, the default neural network formation configuration corresponds to a neural network formation configuration that handles a wider variety of inputs, such as a variety of inputs across more operating conditions. In other words, relative to the neural network formation configuration for the particular operating condition, the deep neural network configured using the default neural network formation configuration handles a wider variety of communications.

[0144] At 1325, UE 110 forms a deep neural network using the neural network formation configuration. As an example, the UE extracts the index value transmitted by base station 121 and obtains the neural network formation configuration and / or the neural network formation configuration element by accessing the neural network table using the index value. Alternatively or additionally, UE 110 extracts the processing assignment and forms the deep neural network in the processing chain as specified by the processing assignment.

[0145] At 1330, base station 121 transmits a communication, such as a downlink data channel communication, to UE 110. At 1335, UE 110 processes the communication using the deep neural network. For example, UE 110 processes the downlink data channel communication using the deep neural network to recover the data. As another example, processing the communication includes processing a reply to the communication, wherein UE 110 processes an uplink communication in reply to the downlink communication using the deep neural network.

[0146] Fig.14 The signaling and control transaction diagram 1400 illustrates a second example of signaling and control transactions for communicating a neural network formation configuration. As shown, at 1405, the core network server 302 maintains a neural network table. For example, the core network neural network manager 312 and / or the training module 314 of the core network server 302 generates and / or maintains the neural network table 316 using any combination of mathematically generated training data, data extracted from real-world communications, files, etc. In various implementations, the core neural network server 302 maintains a plurality of neural network tables, wherein each neural network table includes a plurality of neural network formation configuration elements for a specified processing assignment (e.g., a first neural network table specified for data channel communications, a second neural network table specified for control channel communications).

[0147] At 1410, the core network server 302, such as by using Figure 3 The core network interface 320 communicates the neural network table to the base station 121. In some implementations, the core network server communicates multiple neural network tables to the base station. At 1415, the base station 121 communicates the neural network table to the UE 110, such as by transmitting the neural network table using layer 3 messaging, transmitting the neural network table using a downlink control channel, by broadcasting the neural network table, transmitting the neural network table using a UE-specific message, etc.

[0148] At 1420, the core network server 302 selects a neural network formation configuration. As an example, the core network server 302 compares the current operating environment to input characteristics stored within the neural network table and identifies a stored input characteristic that is aligned with the current operating environment (e.g., one or more of channel conditions, UE capabilities, BS capabilities, metrics). The core network server then obtains an index value for the aligned input characteristic, which in turn provides an index value for the neural network formation configuration and / or a neural network formation configuration element. Then, at 1425, the core network server 302 communicates the selected neural network formation configuration to the base station, such as by communicating the index value using the core network interface 320. In some implementations, the core network server communicates the processing assignment along with the neural network formation configuration.

[0149] At 1430, base station 121 forwards the neural network formation configuration to UE 110. As an example, base station 121 transmits an index value to UE 110, such as via layer 2 messaging (e.g., RLC message, MAC control element), to guide the UE to use the neural network formation configuration to form a deep neural network, such as in Fig.13 In an implementation, base station 121 additionally forms a complementary deep neural network based on a neural network formation configuration (not shown here), such as in Figure 6 In Figure 7 725 of Figure 8 845 and / or Figure 9-1 Sometimes, the base station communicates the processing assignment received from the core network server together with the index value. In response to receiving the neural network formation configuration, UE 110 forms a deep neural network at 1435 and uses the deep neural network to process communications received from base station 121 at 1440, such as in Figure 7 730 of Figure 8 810 of FIG. 9 , at 935 of FIG. 9 , and / or Fig.13 Those described in 1335.

[0150] In some implementations, the UE derives the first neural network forming configuration from the second neural network forming configuration when similarities exist in the wireless communication system. To illustrate, consider the example of quasi-corresponding channels. Quasi-corresponding channels are channels in the wireless communication system that have shared or identical properties, such as the same delay spread, the same Doppler spread, the same spatial signature, the same spatial beam, the same spatial direction, the same data rate, etc. Alternatively or additionally, the quasi-corresponding channels have related physical properties that are within a threshold range or value. In various implementations, the UE derives the first neural network forming configuration from the second neural network forming configuration in response to identifying these similarities.

[0151] To illustrate, now consider Fig.15 A third example of signaling and control transactions for conveying a neural network formation configuration is shown in FIG. 1500 . In the third example, the UE infers the neural network formation configuration based on the identified similarities. Figure 1 base station 121 and UE 110, but alternative implementations include a core network server (such as Figure 3 core network server 302).

[0152] At 1505, the base station 121 identifies two or more channels as quasi-corresponding channels. For example, the base station 121 or the core network server 302 (using Figure 2 The base station neural network manager 268 or Figure 3 The core network neural network manager 312 of the base station 121 compares various beam properties of two or more channels, such as direction, strength, divergence, profile, quality, etc. When any number of properties match and / or correlate with each other within a predetermined threshold, the base station 121 determines that the two or more channels are quasi-corresponding.

[0153] At 1510, the base station 121 transmits an indication of quasi-corresponding channels to the UE 110, such as by transmitting an indication of quasi-corresponding channels using layer 2 messaging, layer 3 messaging, and / or layer 1 signaling. The indication indicates that any number of channels are quasi-corresponding to each other, where the channels can be physical channels and / or logical channels. As an example, the indication indicates that a downlink control channel quasi-corresponds to a downlink data channel when the channels have similar data rates. As another example, the indication indicates that two physical channels at different carrier frequencies based on physical channels having similar spatial beams are quasi-corresponding.

[0154] At 1515, the base station 121 transmits an indication of a neural network formation configuration for one of the quasi-corresponding channels. For example, the base station transmits an index value and a processing assignment that specifies the formed deep neural network for processing a first channel (such as a downlink control channel).

[0155] At 1520, UE 110 forms a first deep neural network using the neural network formation configuration for a first channel (e.g., a downlink control channel). At 1525, UE 110 identifies the first channel as one of the quasi-corresponding channels. For example, the UE compares the first channel with the quasi-corresponding channels identified by and received from base station 121. In response to identifying the first channel as one of the quasi-corresponding channels, the UE determines to apply the neural network formation configuration to other channels that quasi-correspond to the first channel. Therefore, at 1530, UE 110 forms a deep neural network for the other quasi-corresponding channels using the neural network formation configuration. By identifying the channel as a quasi-corresponding channel, only the neural network formation configuration needs to be changed for one of the quasi-corresponding channels, and this will result in changing the neural network formation configuration for all quasi-corresponding channels without having to individually change the neural network formation configuration for each other quasi-corresponding channel.

[0156] Having described signaling and control transactions that can be used to communicate a neural network formation configuration, consider now some example methods that can be used to communicate a neural network formation configuration in accordance with one or more implementations. Example Method

[0157] One or more aspects of the configuration are formed according to the conveying neural network. Fig.16 and Fig.17 Describe example methods 1600 and 1700. The order in which the method boxes are described is not intended to be construed as limiting, and any number of the method boxes may be skipped or combined in any order to implement a method or an alternative method. In general, any components, modules, methods, and operations described herein may be implemented using software, firmware, hardware (e.g., fixed logic circuits), manual processing, or any combination thereof. Some operations of the example methods may be described in the general context of executable instructions stored on a computer-readable storage memory locally and / or remotely from a computer processing system, and implementations may include software applications, programs, functions, and the like. Alternatively or additionally, any functions described herein may be performed at least in part by one or more hardware logic components, such as, but not limited to, a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), an application specific standard product (ASSP), a system on a chip (SoC), a complex programmable logic device (CPLD), and the like.

[0158] Fig.16 An example method 1600 for communicating a neural network formation configuration for processing communications transmitted over a wireless communication system is illustrated. In some implementations, the operations of the method 1600 are performed by a network entity, such as one of the base station 120 and / or the core network server 302.

[0159] At 1605, the network entity transmits a neural network table including a plurality of neural network formation configuration elements, wherein each of the plurality of neural network formation configuration elements configures at least a portion of a deep neural network for processing communications transmitted over a wireless communication system. For example, the network entity (e.g., base station 121) transmits the neural network table (e.g., neural network table 272, neural network table 316) to a UE (e.g., UE 110) using a broadcast or multicast message to a group of UEs. As another example, the network entity (e.g., base station 121) transmits the neural network table to a UE (e.g., UE 110) using a UE-specific message. In some implementations, the network entity (e.g., core network server 302) communicates the neural network table to a base station (e.g., base station 121) and directs the base station to transmit the neural network table to the UE (e.g., UE 110). Sometimes, the network entity transmits the neural network table using layer 3 messaging. Alternatively or additionally, the network entity transmits a plurality of neural network tables, wherein each neural network table has a specified processing assignment, as further described.

[0160] At 1610, the network entity selects one or more neural network formation configuration elements from a plurality of neural network formation configuration elements to create a neural network formation configuration. For example, the network entity (e.g., base station 121, core network server 302) selects the neural network formation configuration element by comparing the current operating environment with input characteristics stored in a neural network table and selecting the neural network formation configuration element by correlating or matching the input characteristics with the current operating environment. In some implementations, the network entity selects a set of candidate neural network formation configurations.

[0161] At 1615, the network entity transmits an indication to the user equipment to direct the user equipment to form a deep neural network using the neural network formation configuration and to process communications using the deep neural network. As one example, the network entity (e.g., base station 121) determines an index value of a neural network table corresponding to the neural network formation configuration and / or a neural network formation configuration element set, and transmits the index value to UE 110, such as by transmitting the index value using a downlink control channel, by transmitting the index value in a layer 2 message, etc. As another example, the network entity (e.g., core network server 302) transmits an indication to a base station (e.g., base station 121), and directs the base station to transmit the indication to the user equipment (e.g., UE 110). The network entity alternatively or additionally indicates a processing assignment for a deep neural network formed using the neural network formation configuration, such as by indicating a processing assignment directing the user equipment to process downlink control channel communications with the deep neural network.

[0162] When transmitting the index value, the network entity sometimes specifies rules regarding when to process communications with the corresponding deep neural network. As an example, the network entity (e.g., base station 121, core network server 302) determines a time threshold between data channel communications and control channel communications. When transmitting the index value to guide a user device (e.g., UE 110) to form a deep neural network, the network entity transmits the time value threshold and a rule that guides the user device to use a default formation configuration to form the deep neural network when the timing between the data channel communication and the control channel communication is below the time threshold. Additionally or alternatively, when the data channel and the control channel are within the time threshold and therefore there is not enough time to switch between different DNNs, the rules can guide the user device to use the same neural network formation configuration for data channel communications and control channel communications.

[0163] Fig.17 An example method 1700 is illustrated for communicating a neural network formation configuration for processing communications transmitted over a wireless communication system. In some implementations, the operations of the method 1700 are performed by a user equipment, such as the UE 110.

[0164] At 1705, a user device receives a neural network table including a plurality of neural network forming configuration elements that provide the user device with the ability to configure a deep neural network for processing communications transmitted over a wireless communication system. For example, a user device (e.g., UE 110) receives a neural network table (e.g., neural network table 272) from a base station (e.g., base station 121) in a layer 3 messaging. As another example, the user device receives the neural network table in a multicast or broadcast message. Alternatively or additionally, the user device receives the neural network table in a UE-specific message. In some cases, the user device receives multiple neural network tables, each of which has a specified processing assignment.

[0165] At 1710, a user device receives a message directing the user device to form a deep neural network using a neural network formation configuration based on one or more neural network formation configuration elements in a plurality of neural network formation configuration elements. For example, a user device (e.g., UE 110) receives in the message a set of index values ​​of a neural network table, wherein the set of index values ​​corresponds to a set of neural network formation configuration elements, such as in Fig.13 1325 of Fig.14 1435 and / or Fig.15 In an implementation, a user equipment (e.g., UE 110) receives a downlink control channel message including an index value corresponding to one or more entries in a neural network table. In some implementations, the message includes an indication of a processing assignment for a deep neural network formed using the neural network formation configuration, such as a communication channel processing assignment, a processing chain processing assignment, etc.

[0166] At 1715, the user device forms a deep neural network using the neural network formation configuration by accessing a neural network table to obtain a neural network formation configuration element. In one example, the user device (e.g., UE 110) uses an index value to access the neural network table. Alternatively or additionally, in some implementations, when forming the deep neural network, the user device (e.g., UE 110) determines to use a first deep neural network formed using the neural network formation configuration to process a first communication channel, and identifies a second communication channel that quasi-corresponds to the first communication channel. In response to identifying that the second communication channel is quasi-corresponding, the user device (e.g., UE 110) determines to use a second deep neural network formed using the neural network formation configuration to process the second communication channel.

[0167] In response to forming the deep neural network, the user equipment uses the deep neural network to process communications transmitted over the wireless communication system at 1720. For example, the user equipment (e.g., UE 110) uses the deep neural network to process a downlink communication channel.

[0168] Having described example methods that can be used to communicate neural network formation configurations, consider now a discussion of E2E ML for wireless networks in accordance with one or more implementations. E2E ML for Wireless Networks

[0169] Aspects of end-to-end communication (E2E communication) involve two endpoints exchanging information over a communication path, such as over a wireless network. Sometimes, E2E communication performs a one-way information exchange, where a first endpoint sends information and a second endpoint receives information. At other times, E2E communication performs a two-way information exchange, where two endpoints send and receive information. Endpoints of E2E communication can include any entity that can consume and / or generate information, such as a computing device, application, service, etc. To illustrate, consider an example in which an application executing at a UE exchanges information with a remote service over a wireless network. In this example, the E2E communication corresponds to a communication path between the application and the remote service, where the application and the remote service act as endpoints.

[0170] Although E2E communication involves endpoints that exchange information, E2E communication may alternatively or additionally include intermediate entities (e.g., devices, applications, services) that participate in the information exchange. To illustrate, consider again the example of E2E communication established over a wireless network, where an application at a UE serves as a first endpoint and a remote service serves as a second endpoint. When establishing E2E communication between endpoints, the wireless network utilizes any combination of UEs, base stations, core network servers, remote networks, remote services, etc., such as reference Figure 1 Thus, intermediate entities such as base station 120 and core network 150 participate in establishing E2E communications and / or participate in E2E communications to enable information exchange between endpoints.

[0171] Various factors affect the operational efficiency of E2E communications and how network elements process information exchanged via E2E communications. For example, with reference to E2E communications established using a wireless network, the current operating environment (e.g., current channel conditions, UE location, UE mobility, UE capabilities) affects how accurately the receiving endpoint recovers information (e.g., bit error rate, packet loss). As an example, E2E communications implemented using 5G mmW technology become more susceptible to signal distortion relative to lower frequency sub-6 GHz signals, as further described.

[0172] As another example, various implementations divide wireless network resources differently based on end-to-end analysis of E2E communication, where wireless network resources include any combination of physical hardware, physical spectrum, logical channels, network functions, services provided, quality of service, latency, etc. as examples and not limitations. Wireless network resource partitioning allows the wireless network to dynamically allocate wireless network resources based on expected usage to improve the efficiency of how wireless network resources are used (e.g., reduce the occurrence of unused resources and / or wasted resources). For illustration, consider various devices connected to a wireless network, where the devices have different performance requirements relative to each other (e.g., a first device has a secure data transmission requirement, a second device has a high priority / low latency data transmission requirement, and a third device has a high data rate requirement). For at least some devices, a fixed and / or static distribution of wireless network resources (e.g., a fixed configuration of wireless network resources for implementing E2E communication) can result in unused resources and / or failure to meet the performance requirements of some services. Therefore, partitioning wireless network resources can improve the overall efficiency of how wireless network resources are utilized. However, the partitioning results in changes in how a pair of E2E endpoints exchange information relative to a second pair of E2E endpoints.

[0173] For further demonstration, consider a Quality of Service flow (QoS flow) corresponding to information exchanged in a wireless network. In various implementations, the E2E communication includes and / or corresponds to a QoS flow. Some wireless networks configure QoS flows with operating rules, priority levels, classifications, etc. that affect how information is exchanged via a QoS flow. For example, a QoS profile indicates to the wireless network QoS parameters and / or QoS characteristics for a particular QoS flow, such as a Guaranteed Flow Bit Rate (GFBR) parameter indicating an uplink and / or downlink guaranteed bit rate for the QoS flow, a Maximum Flow Bit Rate (MFBR) parameter indicating a maximum uplink and / or downlink bit rate for the QoS flow, an Allocation and Retention Priority (ARP) parameter indicating a priority level, preemption capability, and / or preemption vulnerability for the QoS flow, a Reflected QoS Attribute (RQA) indicating that the type of traffic carried on the QoS flow is subject to reflected QoS (e.g., implicit update), a Notification Control parameter indicating whether notification is requested when a guaranteed flow bit rate cannot be guaranteed, or a notification control parameter indicating a collective non-guaranteed bit rate (Non-GBR) flow associated with a particular UE for QoS flow recovery. The parameters and / or characteristics of the configuration of a specified QoS flow may include an aggregate bit rate parameter for the periodic aggregate bit rate, default parameters for 5QI and ARP priority levels, a maximum packet loss rate (MPLR) for uplink and / or downlink indicating the maximum rate of lost packets for a QoS flow, a resource type characteristic indicating the type of resources that can be used by the QoS flow (e.g., GBR resource type, delay critical GBR resource type, non-GBR resource type), a scheduling priority level characteristic that distinguishes multiple QoS flows of the same UE, a packet delay budget characteristic that provides an upper limit on how long a packet can be delayed, a packet error rate characteristic that indicates an upper limit on the rate at which PDUs are not successfully received, an average window characteristic that indicates a data window for calculating GFBR and / or MFBR, a maximum data burst characteristic that indicates the maximum amount of data that needs to be served over a predefined time period, etc. In some implementations, the parameters and / or characteristics that specify the configuration of a QoS flow can be preconfigured (e.g., by default) and / or dynamically communicated, such as through a QoS profile. These changes affect how the wireless network divides various wireless network resources to support QoS flow configurations.

[0174] Fig.18 An example environment 1800 is illustrated that illustrates example changes between exemplary QoS flows according to one or more implementations. The environment 1800 includes Figure 1 UE 110 and base station 120, and Figure 3The core network server 302 of the wireless network system. In the environment 1800, the core network server 302 implements aspects of the user plane function 1802 (UPF 1802), but it should be understood that the UPF can be implemented using alternative or additional devices. The UPF 1802 provides various services and / or features to the wireless network system, such as packet forwarding, packet routing, interconnection with a data network (e.g., the Internet), data buffering, packet inspection, QoS processing, policy enforcement, etc.

[0175] In environment 1800, UE 110 establishes multiple QoS flows through UPF 1802 to obtain access to the data network. For example, UE 110 includes three applications: application 1804, application 1806, and application 1808, each of which has different performance requirements (e.g., resource type, priority level, packet delay budget, packet error rate, maximum data burst, average window, security level). These different performance requirements cause the wireless network to divide the wireless network resources assigned to the corresponding QoS flows differently from each other. For example, consider a scenario in which application 1804 corresponds to a game application, application 1806 corresponds to an augmented reality application, and application 1808 corresponds to a social media application. In some instances, the game application interacts with a remote service (through a data network) to connect with another game application to exchange audio in real time, exchange video in real time, exchange commands, exchange views, etc., so that the game application has high data volume and low latency performance requirements. The augmented reality application also interacts with the remote service through the data network to transmit location information, and then receives image data overlaid on the camera image generated at UE 110. Compared to gaming applications, augmented reality applications utilize less data, but have a certain time sensitivity to maintain synchronization between the current location and the corresponding image overlay. Finally, social media applications interact with remote services through data networks to receive feed information, wherein the feed information has less data volume and time criticality than the data consumed by augmented reality applications and / or gaming applications.

[0176] Based on these performance requirements, the wireless network establishes QoS flows between the application and the data network (through the UPF), wherein each QoS flow has a different configuration associated with partitioning the wireless network resources. For example, QoS flow 1810 connects application 1804 to the data network through UPF 1802, wherein the wireless network constructs the QoS flow based on QoS requirements, QoS parameters, and / or QoS characteristics (e.g., resource type, priority level, packet delay budget, packet error rate, maximum data burst, average window, security level), which QoS requirements, QoS parameters, and / or QoS characteristics indicate high data volume performance requirements and time sensitivity performance requirements. In an implementation, the QoS requirements, QoS parameters, and / or QoS characteristics included in the QoS profile correspond to the performance requirements of the QoS flow.

[0177] To illustrate, the wireless network (e.g., base station 120, core network server 302) processes a QoS profile associated with a QoS flow 1810, which configures any combination of GFBR parameters, maximum data burst characteristics, ARP parameters, etc. based on the performance requirements of the QoS flow 1810. The wireless network then constructs the QoS flow by partitioning wireless network resources based on the QoS parameters and / or characteristics, where the QoS flow 1810 connects the application 1804 to the data network through the UPF 1802. Similarly, the QoS flow 1812 connects the application 1806 to the data network through the UPF 1802, where the wireless network partitions the wireless network resources used by the QoS flow based on the performance requirements associated with the application 1806. Finally, the QoS flow 1814 connects the application 1808 to the data network (through the UPF 1802), where the wireless network partitions the wireless network resources used by the QoS flow based on the performance requirements associated with the application 1808, such as by processing a QoS profile where the QoS parameters and / or characteristics reflect the smaller data volume and less time-critical requirements of the QoS flow 1814. In various implementations, UPF 1802 acts as an edge of the wireless network because UPF 1802 provides connectivity to data networks outside the wireless network. Thus, within the wireless network, application 1804 and UPF 1802 act as endpoints for E2E communications based on QoS flow 1810, application 1806 and UPF 1802 act as endpoints for E2E communications based on QoS flow 1812, and application 1808 and UPF 1802 act as endpoints for E2E communications based on QoS flow 1814.

[0178] Environment 1800 illustrates an overview of an example architecture for a wireless network to construct QoS flows. In the example architecture, the wireless network constructs a protocol data unit session (PDU session 1816) that supplies data connectivity between UEs 110 and to an external data network through UPF 1802. In an implementation, a PDU session transmits data based on a specific service type (e.g., Internet Protocol (IP) service, Ethernet type service, unstructured type service, etc.). In this example, QoS flow 1810, QoS flow 1812, and QoS flow 1814 use the same service type and belong to a PDU session 1816 corresponding to transmitting data associated with the service type.

[0179] To support data connectivity, data radio bearer 1818 and data radio bearer 1820 provide a radio interface connecting UE 110 and base station 120. In some implementations, data radio bearer 1818 and data radio bearer 1820 communicate information over the radio interface using different technologies, resource allocations, etc. Thus, the example architecture maps each QoS flow to a data radio bearer with a configuration that supports the corresponding performance requirements of the QoS flow. In this example, the wireless network maps data radio bearer 1818 to QoS flow 1810 and QoS flow 1812, and maps data radio bearer 1820 to QoS flow 1814. Thus, the data radio bearer can support multiple QoS flows.

[0180] Tunnel 1822 corresponds to non-radio interface functions associated with QoS flows provided by UPF 1802, such as packet forwarding. In an implementation, tunnel 1822 corresponds to a tunneling protocol for transmitting data between entities. Thus, base station 120 uses tunnel 1822 to communicate information exchanged on different QoS flows to the data network through UPF 1802. In an implementation, tunnel 1822 provides the wireless network with the ability to exchange services associated with PDU session 1816.

[0181] While the configurability of QoS flows provides wireless networks with the flexibility to dynamically modify how wireless network resources are allocated, the configurability increases the complexity of how the wireless network processes information exchanged between endpoints. Some implementations train DNNs to perform some or all of the complex processing associated with exchanging information using E2E communications with various configurations. By training the DNNs on different processing chain operations and / or wireless network resource partitions, the DNNs can replace traditional complex functions, as further described. The use of DNNs in E2E communications also allows network entities to adapt the DNNs to changing operating conditions, such as by modifying various parameter configurations (e.g., coefficients, layer connections, kernel size).

[0182] One or more implementations determine an E2E ML configuration (E2EML configuration) for processing information exchanged via E2E communication. For example, referring to Fig.18 , some implementations determine a first E2E ML configuration for processing communications exchanged using QoS flow 1810, a second E2E ML configuration for processing communications exchanged using QoS flow 1812, and a third E2E ML configuration for processing communications exchanged using QoS flow 1814. In some cases, an end-to-end machine learning controller (E2EML controller) obtains capabilities of devices associated with end-to-end communications in a wireless network, such as machine learning (ML) capabilities of devices participating in the E2E communications, and determines the E2E ML configuration based on the ML capabilities of the devices (e.g., supported ML architectures, supported number of layers, available processing power, memory limitations, available power budget, fixed-point processing versus floating-point processing, maximum kernel size capabilities, compute capabilities). Alternatively or additionally, the E2E ML controller identifies a current operating environment and determines the E2E ML configuration based on the current operating environment. In determining the E2E ML configuration, some implementations of the E2E ML controller partition the E2E ML configuration based on the devices participating in the E2E communication, and communicate the respective partition of the E2E ML configuration to each respective device.

[0183] For demonstration purposes, consider Fig.19 , which illustrates an example environment 1900 in which an example E2E ML configuration for E2E communication includes a DNN operating at multiple devices. The environment 1900 includes Figure 1 UE 110, base station 120 and remote service 170 and Figure 3 1900 . In an implementation, UE 110 and remote service 170 use E2E communication 1902 and E2E communication 1904 to exchange information with each other. For clarity, E2E communication 1902 and 1904 are illustrated as separate unidirectional E2E communications, such that the information exchange on each communication E2E corresponds to one direction (e.g., E2E communication 1902 includes uplink transmissions and E2E communication 1904 includes downlink transmissions), but it should be understood that in alternative or additional implementations, E2E communication 1902 and E2E communication 1904 correspond to unidirectional and bidirectional E2E communications including both of which are represented by dashed line 1906 in environment 1900. In an implementation, E2E communication 1902 and E2E communication 1904 (in combination) correspond to a single QoS flow, such as Fig.18 QoS flow 1810, QoS flow 1812 or QoS flow 1814, therefore, Fig.18 Each of the multiple QoS flows illustrated in FIG. 1 has a corresponding E2E ML configuration.

[0184] The environment 1900 also includes an E2E ML controller 318 implemented by the core network server 302, wherein the E2E ML controller 318 determines an E2E ML configuration for the E2E communication 1902 and / or the E2E communication 1904. In some implementations, such as when each E2E communication corresponds to a unidirectional information exchange, the E2E ML controller determines a first E2E ML configuration for the E2E communication 1902 and a second E2E ML configuration for the E2E communication 1904. In other implementations, the E2E ML controller determines an E2E ML configuration for bidirectional E2E communication including both the E2E communications 1902 and 1904. For example, in response to UE 110 requesting a connection to remote server 170, such as through a call by an application, the E2E ML controller determines an E2E ML configuration for the corresponding connection based on any combination of ML capabilities of UE 110 (e.g., supported ML architecture, number of layers supported, processing power available for ML processing, memory constraints applied to ML processing, power budget available for ML processing, fixed-point processing versus floating-point processing), performance requirements associated with the requested connection (e.g., resource type, priority registration, packet delay budget, packet error rate, maximum data burst size, average window, security level), available wireless network resources, ML capabilities of intermediate devices (e.g., base station 120, core network server 302), current operating environment (e.g., channel conditions, UE location), etc. As an example, refer to Figure 9-1 , Figure 9-2 and Fig.10 , the E2E ML controller 318 receives base station metrics and / or UE metrics describing the current operating environment through the core network server 302. As another example, refer to Fig. 22 , the E2E ML controller receives the base station ML capabilities and / or UE capabilities through the core network server 302 .

[0185] In one or more implementations, the E2E ML controller 318 analyzes the neural network table to determine the E2E ML configuration based on any combination of device capabilities, wireless network resource partitioning, operating parameters, current operating environment, ML capabilities, etc. Although described as being implemented by the core network server 302, in alternative or additional implementations, the E2E ML controller 318 may be implemented by another network entity, such as the base station 120.

[0186] For illustration, and with reference to Fig.12, the training module 270 and / or the training module 314 trains the machine learning module 400 with variations of the training data 1202 that reflect different combinations of capabilities, wireless network resource partitions, operating parameters, current operating environments, etc. The training module extracts the architecture and / or parameter configuration (e.g., the architecture and / or parameter configuration 1208) and stores it in the neural network table, so that at a later point in time, the E2E ML controller 318 accesses the neural network table to obtain and / or identify the neural network formation configuration corresponding to the determined E2E ML configuration. The E2E ML controller 318 then communicates the neural network formation configuration to various devices and directs the respective devices to form respective DNNs, as further described.

[0187] When determining the E2E ML configuration, the E2E ML controller 318 sometimes divides the E2E ML configuration based on the devices participating in the corresponding E2E communication. For example, the E2E ML controller 318 determines a first partition of the E2E ML configuration corresponding to the processing information at the UE 110, a second partition of the E2E ML configuration corresponding to the processing information at the base station 120, and a third partition of the E2E ML configuration corresponding to the processing information at the core network server 302, wherein the determination of the partition can be based on any combination of capabilities, wireless network resource divisions, operating parameters, current operating environment, etc.

[0188] As an example, consider an E2E communication corresponding to voice transmission over a wireless network, such as E2E communication 1902, E2E communication 1904, and / or a combination of these two E2E communications. In determining an E2E ML configuration for the E2E communication, the E2E ML controller 318 may alternatively or additionally identify a performance requirement for the E2E communication indicating a large amount of data transmission with a low latency requirement. Based on the performance requirement, the E2E ML controller identifies an E2E ML configuration that, when formed by a corresponding DNN, performs end-to-end functions of exchanging voice communications and meets the performance requirement. For illustration, the E2E ML controller determines an E2E ML configuration that performs end-to-end functions for transmitting voice from a UE to a core network server (such as signal processing, voice encoding, channel encoding, and / or channel modulation on the UE side, channel decoding, demodulation, and / or signal processing on the base station side, decoding voice on the core network server side, etc.), and selects a configuration designed to meet the performance requirement.

[0189] Some implementations partition the E2E ML configuration based on the ML capabilities and / or performance requirements of the devices participating in the E2E communication. For example, the UE may have fewer processing resources (e.g., processing power, memory constraints, quantization constraints, fixed-point versus floating-point calculations, FLOPS, power availability relative to the base station and / or core network server), which may be indicated by the ML capabilities. In response to identifying different processing resources through analysis of the ML capabilities, the E2E ML controller partitions the E2E ML configuration so that a first partition (e.g., at the UE 110) forms a DNN that performs less processing than a DNN formed by a second or third partition (e.g., at the base station, at the core network server). Alternatively or additionally, the E2E ML controller partitions the E2E ML configuration to produce a neural network designed not to exceed the device capabilities. For example, based on analyzing the capabilities, the E2E ML controller directs the UE to form a DNN with fewer layers and a smaller kernel size relative to a DNN formed by the base station and / or core network server based on the UE's processing constraints. Alternatively or additionally, the E2E ML controller partitions the E2E ML configuration to form a neural network (e.g., convolutional neural network, long short-term memory (LSTM) network, partially connected, fully connected) at the UE having an architecture that processes information without exceeding the memory constraints of the UE. In some examples, the E2E ML controller calculates whether the amount of computation performed at each device collectively meets a performance requirement corresponding to a delay budget, and determines an E2E ML configuration designed to meet the performance requirement.

[0190] In the environment 1900, the E2E ML controller 318 determines a first E2E ML configuration for processing information exchanged via the E2E communication 1902, and determines to partition the first E2E ML configuration across multiple devices, such as by partitioning the first E2E ML configuration across the UE 110, the base station 120, and the core network server 302 based on device capabilities. In other words, some implementations determine an E2E ML configuration corresponding to a distributed DNN in which multiple devices implement and / or form parts of the DNN. To communicate the partition, the E2E ML controller 318 identifies a first neural network formation configuration (NN formation configuration) corresponding to a first partition of the E2E ML configuration, and communicates the first NN formation configuration to the UE 110 by using the core network server 302. The E2E ML controller 318 and / or the core network server 302 then directs the UE to form a user device-side deep neural network 1908 (UE-side DNN 1908) for processing the information exchanged via the E2E communication 1902. Similarly, the E2E ML controller 318 identifies a second NN formation configuration corresponding to the second partition of the E2E ML configuration, and communicates the second NN formation configuration to the base station 120. The E2E ML controller 318 and / or the core network server 302 then directs the base station 120 to use the second NN formation configuration to form a base station-side deep neural network 1910 (BS-side DNN 1910) for processing information exchanged via the E2E communication 1902. The E2E ML controller 318 also identifies a third NN formation configuration and communicates it to the core network server 302 for use in forming a core network server-side deep neural network 1912 (CNS-side DNN 1912) for processing information exchanged via the E2E communication 1902.

[0191] In an implementation, the E2E ML controller 318 partitions the E2E ML configuration to distribute processing computations performed on the E2E communication such that the UE-side DNN 1908 performs less processing relative to the BS-side DNN 1910 (e.g., the UE-side DNN 1908 uses fewer layers, fewer data processing points, etc. relative to the BS-side DNN 1910). Alternatively or additionally, the E2E ML controller 318 partitions the E2E ML configuration such that the BS-side DNN 1910 performs less processing relative to the CNS-side DNN 1912. In combination, the processing performed by the UE-side DNN 1908, the BS-side DNN 1910, and the CNS-side DNN 1912 exchange information across the E2E communication 1902.

[0192] In a similar manner, the E2E ML controller 318 determines a second E2E ML configuration for processing information exchanged via the E2E communication 1904, wherein the E2E ML controller partitions the second E2E ML configuration across multiple devices. In the environment 1900, the partition corresponds to a core network server-side deep neural network 1914 (CNS-side DNN 1914), a base station-side deep neural network 1916 (BS-side DNN 1916), and a user device-side deep neural network 1918 (UE-side DNN 1918). In combination, the processing performed by the CNS-side DNN 1914, the BS-side DNN 1916, and the UE-side DNN 1918 corresponds to exchanging information using the E2E communication 1904. Although the E2E ML controller separately determines the first and second E2E ML configurations for unidirectional E2E communications (e.g., E2E communications 1902 and 1904) in environment 1900, it should be understood that in alternative or additional implementations, the E2E ML controller 318 determines a single E2E ML configuration corresponding to exchanging bidirectional information using E2E communications. For example, the E2E controller 318 may determine Fig.18 1810 , a second (bidirectional) E2E ML configuration for QoS flow 1812 , and a third (bidirectional) E2E ML configuration for QoS flow 1814 .

[0193] Fig. 20 Another example environment 2000 is illustrated in which an E2E ML configuration for E2E communication includes a DNN operating across multiple devices. The environment 2000 includes a first instance of UE 110 (designated as UE 2002), Figure 1 A first instance of base station 120 (designated as base station 2004), Figure 3 Core network server 302, Figure 1 A second instance of base station 120 (designated as base station 2006) and a second instance of UE 110 (designated as UE 2008).

[0194] In environment 2000, core network server 302 operates in a wireless network (via base station 2004 and base station 2006) that exchanges information between UE 2002 and UE 2008. E2E communication 2010 represents a one-way E2E communication between a first endpoint (e.g., UE 2002) and a second endpoint (e.g., UE 2008), and E2E communication 2012 represents a one-way E2E communication between UE 2008 and UE 2002. However, as indicated by dashed line 2014, E2E communication 2010 and E2E communication 2012 may collectively represent two-way E2E communication, as further described.

[0195] Similar to reference Fig.19 As described above, the E2E ML controller 318 determines a first E2E ML configuration for processing (unidirectional) communications exchanged using E2E communications 2010, a second E2E ML configuration for processing (unidirectional) communications exchanged using E2E communications 2012, and / or an E2E ML communication for processing (bidirectional) communications exchanged using E2E communications 2010 and E2E communications 2012. In determining the E2E ML configuration, the E2E ML controller 318 analyzes any combination of device capabilities, wireless network resource partitioning, operating parameters, current operating environment, performance requirements, etc. Alternatively or additionally, the E2E ML controller 318 determines partitioning of the E2E ML configuration to distribute processing functions associated with the E2E ML configuration across multiple devices.

[0196] To illustrate, in order to process the information exchanged through E2E communication 2010, UE 2002 receives a first neural network formation configuration from E2E ML controller 318 through core network server 302, the first neural network formation configuration corresponding to the first partition of the E2E ML configuration and mapped to an entry in a neural network table that can be used to form a user equipment side deep neural network 2016 (UE side DNN 2016). Base station 2004 receives a second neural network formation configuration corresponding to the second partition of the E2E ML configuration and can be used to form a base station side deep neural network 2018 (BS side DNN 2018). In a similar manner, core network server 302 forms a core network side deep neural network 2020 (CNS side DNN 2020) based on a third neural network formation configuration, base station 2006 forms a base station side deep neural network 2022 (BS side DNN 2022) based on a fourth neural network formation configuration, and UE 2008 forms a user equipment side deep neural network 2024 (UE side DNN 2024) based on a fifth neural network formation configuration. Therefore, in environment 2000, UE-side DNN 2016, BS-side DNN 2018, CNS-side DNN 2020, BS-side DNN 2022, and UE-side DNN 2024 correspond to the divided E2E ML configuration determined by the E2E ML controller 318, and jointly perform processing for exchanging information from UE 2002 to UE 2008 using E2E communication 2010.

[0197] Regarding the E2E communication 2012, the E2E ML controller 318 determines the partitioned E2E ML configuration and communicates the corresponding parts of the partitioned E2E ML configuration to the devices participating in the E2E communication 2012. In order to exchange information using the E2E communication 2012, the UE 2008 forms a user equipment side deep neural network 2026 (UE side DNN 2026), the base station 2006 forms a base station side deep neural network 2028 (BS side DNN 2028), the core network server 302 forms a core network side deep neural network 2030 (CNS side DNN 2030), the base station 2004 forms a base station side deep neural network 2032 (BS side DNN 2032), and the UE 2002 forms a user equipment side deep neural network 2034 (UE side DNN 2034).

[0198] In an implementation, the E2E ML controller 318 periodically re-evaluates metrics, performance requirements, radio link performance, processing capabilities of the device, or other aspects that affect or provide an indication of the current operating environment and / or current performance (e.g., bit error, BLER) to determine whether to update the E2EML configuration. For example, the E2E ML controller 318 determines modifications (e.g., parameter changes) to an existing DNN to better accommodate the performance requirements of devices, applications, and / or transmissions in the wireless network. A UE that changes location may affect radio link performance, or a user opening an application at a UE may reduce the processing capabilities that the user device can provide for machine learning. By re-evaluating dynamically changing conditions (e.g., changes in operating environment, changes in devices), the E2E ML controller is able to modify or update the E2EML configuration to improve the overall efficiency of how wireless network resources are utilized.

[0199] Having described E2E ML for wireless networks, consider now a discussion of signaling and control transactions over wireless communication systems that can be used in various aspects of E2E ML for wireless networks. Signaling and control transactions for E2E ML in wireless networks

[0200] Fig.21 , 22 , 23 and 24 illustrate example signaling and control transaction diagrams between a core network server, a base station and a user equipment according to one or more aspects of E2E ML for wireless networks. The signaling and control transactions may be performed by Figure 1 The base station 120 and the UE 110 or Figure 3 The core web server 302 uses Figure 1-20 components to execute.

[0201] Fig.21Figure 2100 illustrates a first example of signaling and control transactions for an E2E ML for a wireless network. As shown, at 2105, UE 110 requests to establish E2E communication. For example, the request implicitly requests to establish E2E communication by requesting to establish a communication path for an application in response to detecting a call of the application. Example applications include virtual reality applications, augmented reality applications, voice over Internet protocol (VoIP) applications, real-time gaming applications, video streaming applications, audio streaming applications, file transfer applications (e.g., photo / image transfer applications, portable document format (PEF) file transfer applications), social media applications, and other forms of interactive communication. In various implementations, UE 110 explicitly requests to establish E2E communication, such as by requesting to establish a QoS flow with performance requirements, as further described. Alternatively or additionally, UE 110 requests (implicitly or explicitly) to establish E2E communication for other types of communications, such as IoT messaging and / or communications, vehicle-to-everything (V2X) messaging and / or communications, IP Multimedia Core Network Subsystem (IMS) signaling, remote control communications, intelligent transport systems, or discrete automation.

[0202] In some implementations, the UE 110 includes a QoS identifier in the request, wherein the QoS identifier indicates characteristics regarding the requested E2E communication. For illustration, the UE 110 includes a fifth generation quality of service identifier (5QI) or a quality of service class identifier (QCI) in the request, wherein the 5GCI and / or QCI can be analyzed to determine QoS requirements, QoS parameters, QoS characteristics, etc.

[0203] At 2110, base station 120 relays and / or forwards the request to the core network server. In some cases, base station 120 generates a new message encapsulating the request and sends it to core network server 302, while in other cases, base station 120 relays the request from UE 110 to core network server 302.

[0204] At 2115, the core network server 302 determines the E2E ML configuration via the E2E ML controller 318. For example, the E2E ML controller analyzes the performance requirements of the E2E communication, such as the processing requirements associated with establishing the QoS flow 1810, the QoS flow 1812, the QoS flow 1814, the E2E communication 1902 and / or the E2E communication 1904, the E2E communication 2010 and / or the E2E communication 2012, etc. For example, the E2E ML controller 318 analyzes the QoS identifier to determine the performance requirements. Alternatively or additionally, the E2E ML controller analyzes the capabilities of the devices participating in the E2E communication, such as the reference Fig. 22Sometimes, the E2E ML controller analyzes the partitioning of wireless network resources associated with the E2E communication. As another example, the E2E ML controller 318 analyzes metrics characterizing the current operating environment, such as reference Figure 9-1 , Figure 9-2 or Fig. 22 As described, and based on the current operating environment, the E2E ML configuration is determined. In an implementation, the E2E ML configuration corresponds to a unidirectional E2E communication, while in other implementations, the E2E communication corresponds to a bidirectional E2E communication.

[0205] In implementations, determining the E2E ML configuration includes determining any combination of structures and / or parameter configurations that can be used to create a NN formation configuration (e.g., a combination of one or more NN formation configuration elements) that defines and / or forms the DNN. For example, in some implementations, the E2E ML controller determines a first E2E ML configuration that includes an architectural configuration and a parameter configuration for forming the DNN. Alternatively or additionally, the E2E ML controller then determines a second E2E ML configuration that includes parameter changes (and excludes the architectural configuration) for updating the first E2E ML configuration based on the changed operating environment, such as a reference Figure 9-1 , Figure 9-2 and Fig.23 Those described.

[0206] In some implementations, the core network server 302 analyzes the neural network table through the E2E ML controller 318 to obtain one or more neural network formation configurations corresponding to the E2E ML configuration. Fig.19 and 20 The core network server partitions the E2E ML configuration across multiple devices and determines a corresponding neural network formation configuration for each partition, wherein each corresponding neural network formation configuration maps to one or more entries in a neural network table.

[0207] At 2125, the core network server 302 communicates the E2E ML configuration to the device associated with the E2E communication. To illustrate, the core network server communicates the first neural network formation configuration to the first device (e.g., base station 120), communicates the second neural network formation configuration to the second device (e.g., UE 110 via base station 120 at 2130), and so on. This includes identifying the neural network configuration corresponding to forming the DNN at the core network server. In some implementations, the core network server 302 communicates with reference to Figure 7 The described neural network formation configuration is similar to the neural network formation configuration.

[0208] At 2135, core network server 302 forms a DNN (e.g., DNN 1912, DNN 1914, DNN 2020, DNN 2030) based on the E2E ML configuration determined at 2115. In an implementation, the DNN formed by the core network server performs at least some processing for exchanging information using E2E communication. Similarly, at 2140, base station 120 forms a DNN (e.g., DNN 1910, DNN 1916, DNN 2018, DNN 2022, DNN 2028, DNN 2032) based on the E2E ML configuration determined at 2115. For example, base station 120 accesses a neural network table to obtain a DNN such as reference 2140. Fig.12 In an implementation, the DNN formed by base station 120 performs at least some processing for exchanging information using E2E communication. At 2145, UE 110 forms a DNN (e.g., DNN 1908, DNN 1918, DNN 2016, DNN 2024, DNN 2026, DNN 2034) based on the E2E ML configuration determined at 2115. For example, UE 110 accesses a neural network table to obtain a DNN as described in reference Fig.12 In an implementation, the DNN formed by the UE 110 performs at least some processing for exchanging information using E2E communication.

[0209] Then, at 2150, at least the core network server 302, the base station 120, and the UE 110 use the DNN to process communications, wherein the communications correspond to information exchanged via E2E communications. Figure 6 , the DNN at the device processes uplink and / or downlink information.

[0210] Fig. 22 Figure 2200 illustrates a second example of signaling and control transactions for E2E ML for a wireless network. As shown, at 2205, the core network server 302 requests machine learning (ML) capabilities from one or more devices in the wireless network. For example, the E2E ML controller 318 initiates a request for ML capabilities, where the core network server 302 sends the request to the base station 120 via an NG2 interface, an S1 interface, etc. In some implementations, the capability request explicitly indicates a request for ML capabilities, such as by including fields and / or information elements in the request. In other implementations, the capability request implicitly indicates a request for ML capabilities, such as by a capability request without any explicit ML capability fields and / or information elements. Various implementations explicitly or implicitly request ML capabilities as part of constructing E2E communications, such as through a non-access stratum (NAS) message and / or QoS flow used to establish E2E communications.

[0211] At 2210, the base station 120 transmits a request for ML capability to the UE 110, wherein the request explicitly or implicitly requests ML capability. As one example, the base station 120 relays the ML capability request from the core network server 302. As another example, the base station 120 generates and transmits an ML capability request message to the UE 110 in response to receiving the ML capability request from the core network server 302. In various implementations, the base station 120 transmits the explicit or implicit request for ML capability as part of establishing E2E communication.

[0212] At 2215, the base station 120 sends the base station machine learning capabilities (BS ML capabilities) to the core network server 302. Similarly, at 2220, the UE 110 transmits the user equipment machine learning capabilities (UE ML capabilities) to the base station, where the base station 120 relays the UE ML capabilities to the core network server at 2225. Sometimes, the core network server forwards the BS ML capabilities and / or the UE ML capabilities to the E2E ML controller 318. The BS ML capabilities and / or the UE ML capabilities include any combination of capability information, such as supported ML architectures, number of layers supported, available processing power, memory limits, maximum kernel size, available power budget, computing power, and the like.

[0213] At 2230, the core network server 302 determines the E2E ML configuration through the E2E ML controller 318. For example, the E2EML controller analyzes the performance requirements of the E2E communication, such as the processing requirements of the QoS flow 1810, the QoS flow 1812, the QoS flow 1814, the E2E communication 1902 and / or the E2E communication 1904, the E2E communication 2010 and / or the E2E communication 2012, etc. Alternatively or additionally, the E2EML controller analyzes the capabilities of the devices participating in the E2E communication, such as the BS ML capabilities received by the core network server at 2215, the UE ML capabilities received by the core network server at 2225, the processing capabilities, etc. Sometimes, the E2E ML controller analyzes the partitioning of wireless network resources associated with the E2E communication. As another example, the E2E ML controller 318 analyzes metrics characterizing the current operating environment, such as reference Figure 9-1 and Figure 9-2 As described, and based on the current operating environment, the E2E ML configuration is determined. In an implementation, the E2E ML configuration corresponds to a unidirectional E2E communication, while in other implementations, the E2E communication corresponds to a bidirectional E2E communication.

[0214] Sometimes, the core network server 302 analyzes the neural network table through the E2E ML controller 318 to obtain one or more neural network formation configurations corresponding to the E2E ML configuration. Fig.18 , 19 and 20 the core network server partitions the E2E ML configuration across multiple devices and determines a corresponding neural network formation configuration for each partition, wherein each corresponding neural network formation configuration maps to an entry in the neural network table.

[0215] At 2235, the core network server 302 communicates the E2E ML configuration to the device associated with the E2E communication. To illustrate, the core network server communicates the first neural network formation configuration to the first device (e.g., base station 120), communicates the second neural network formation configuration to the second device (e.g., UE 110 via base station 120 at 2240), etc. This includes identifying the neural network configuration corresponding to forming the DNN at the core network server. In some implementations, the core network server 302 communicates the E2E ML configuration to the device associated with the E2E communication. For example, the core network server 302 communicates the first neural network formation configuration to the first device (e.g., base station 120), communicates the second neural network formation configuration to the second device (e.g., UE 110 via base station 120 at 2240), etc. Figure 7 The described approach is similar to the way a neural network formation configuration is communicated.

[0216] At 2245, core network server 302 forms a DNN (e.g., DNN 1912, DNN 1914, DNN 2020, DNN 2030) based on the E2E ML configuration determined at 2230. In an implementation, the DNN formed by the core network server performs at least some processing to exchange information using E2E communication. Similarly, at 2250, base station 120 forms a DNN (e.g., DNN 1910, DNN 1916, DNN 2018, DNN 2022, DNN 2028, DNN 2032) based on the E2E ML configuration determined at 2230. For example, base station 120 accesses a neural network table to obtain a neural network table such as reference 2230. Fig.12 In an implementation, the DNN formed by the base station 120 performs at least some of the processing to exchange information using E2E communication. At 2255, the UE 110 forms a DNN (e.g., DNN 1908, DNN 1918, DNN 2016, DNN 2024, DNN 2026, DNN 2034) based on the E2E ML configuration determined at 2230. For example, the UE 110 accesses the neural network table to obtain the neural network table as shown in the reference Fig.12 In an implementation, the DNN formed by the UE 110 performs at least some processing to exchange information using E2E communication.

[0217] Then, at 2260, at least the core network server 302, the base station 120, and the UE 110 use the DNN to process communications, where the communications correspond to information exchanged via the E2E communication. Figure 6 , the DNN at the device processes uplink and / or downlink information.

[0218] Fig.23 FIG. 2300 illustrates a third example of signaling and control transactions for E2E ML for a wireless network. In an implementation, aspects of FIG. 2300 may be combined with Fig.21 The signaling and control transaction diagram 2100 and / or Fig. 22 The signaling and control transaction diagram 2200 operates as or as a continuation thereof.

[0219] As shown, at 2305, the core network server 302 receives one or more metrics associated with the wireless network. Figure 9-1 and Figure 9-2 , the core network server 302 receives UE metrics from the UE 110 and / or BS metrics from the base station 120 based on the communication in the wireless network (such as uplink received power, uplink SINR, uplink packet error, uplink throughput, timing measurement, power information, SINR information, CQI, CSI, Doppler feedback, QoS, delay, etc.). This may include metrics generated based on information exchanged through E2E communication and / or metrics associated with information exchanged through other communication paths. In an implementation, the UE 110, the base station 120, and the core network server 302 participate in the same E2E communication to exchange information from a first endpoint to a second endpoint of the E2E communication, such as QoS flow 1810, QoS flow 1812, QoS flow 1814, E2E communication 1902 and / or E2E communication 1904, E2E communication 2012 and / or E2E communication 2012, etc.

[0220] At 2310, the core network server 302 determines the E2E ML configuration based on the metrics via the E2E ML controller 318. For illustration, refer to Figure 9-1 and Figure 9-2 , some implementations identify changes in the current operating environment and determine to update a DNN for processing information exchanged via E2E communication, such as by determining parameter changes applied to a DNN architecture. Alternatively or additionally, some implementations determine an initial E2E ML configuration for establishing E2E communication. As further described, determining the E2E ML configuration sometimes includes determining a partitioned E2E ML configuration in which different devices form and use respective DNNs to process information exchanged via E2E communication, such as by determining that a respective neural network of each respective device participating in the E2E communication forms a configuration.

[0221] Determining the E2E ML configuration may include identifying any combination of an architectural configuration and / or a parameter configuration / change of the E2E ML configuration. The determined E2E ML configuration may correspond to a small change to the current E2E ML configuration, wherein the small change updates parameters (e.g., coefficients, weights, filters) of the architectural configuration applied to the DNN. For example, the small change updates the DNN to account for changes in the operating environment identified by analyzing the metrics received at 2305 (e.g., SINR changes, Doppler feedback changes, power level changes, BLER changes). Alternatively or additionally, the determined E2E ML configuration corresponds to a large change that modifies the ML architecture of the DNN by reconfiguring node and / or layer connections based on analyzing the metrics received at 2305, such as a change in power state (e.g., a transition from a radio resource connected state to an idle state).

[0222] At 2315, the core network server communicates the E2E ML configuration. To illustrate, the core network server communicates the partitioned E2E ML configuration by communicating a first neural network formation configuration to a first device (e.g., base station 120), communicating a second neural network formation configuration to a second device (e.g., UE 110 via base station 120 at 2320), etc. Sometimes, the core network server identifies a third neural network configuration corresponding to forming a DNN at the core network server. In some implementations, the core network server 302 communicates a corresponding neural network formation configuration corresponding to a portion of the E2E ML configuration, such as by communicating a mapping to a reference UE 110 as shown in FIG. Figure 7 The index value in the entry of the neural network table.

[0223] At 2325, the core network server 302 forms a DNN (e.g., DNN 1912, DNN 1914, DNN 2020, DNN 2030) based on the E2E ML configuration determined at 2315. For example, referring to Fig.12 , the core network server accesses the neural network table to obtain parameters for forming the DNN. When forming the DNN, the core network server 302 forms a new DNN as part of establishing E2E communication and / or updates an existing DNN for exchanging information using E2E communication. In an implementation, the DNN formed by the core network server corresponds to a portion of an E2E ML configuration, as further described.

[0224] Similarly, at 2330, base station 120 forms a DNN (e.g., DNN 1910, DNN 1916, DNN 2018, DNN 2022, DNN 2028, DNN 2032) based on the E2E ML configuration determined at 2315. For example, referring to Fig.12 , base station 120 accesses the neural network table to obtain parameters for forming the DNN. In forming the DNN, the base station forms a new DNN as part of establishing E2E communication and / or updates an existing DNN for exchanging information using E2E communication. In an implementation, the DNN formed by base station 120 corresponds to a portion of an E2E ML configuration, as further described.

[0225] At 2335, UE 110 forms a DNN (e.g., DNN 1908, DNN 1918, DNN 2016, DNN 2024, DNN 2026, DNN 2034) based on the E2E ML configuration determined at 2310. For example, referring to Fig.12 , UE 110 accesses the neural network table to obtain parameters for forming the DNN. When forming the DNN, UE 110 forms a new DNN as part of establishing E2E communication and / or updates an existing DNN for exchanging information using E2E communication. In an implementation, the DNN formed by UE 110 corresponds to a portion of an E2E ML configuration, as further described.

[0226] Then, at 2340, at least the core network server 302, the base station 120, and the UE 110 use the DNN to process communications, where the communications correspond to information exchanged via the E2E communication. Figure 6 , the DNN at the device processes uplink and / or downlink information. Alternatively or additionally, the DNN at the device processes information exchanged via the QoS flow (e.g., Fig.18 ).

[0227] In an implementation, the core network server 302, the base station 120 and / or the UE 110 iteratively executes Fig.23 The signaling and control transactions depicted in Figure 2300 are represented by dashed lines 2345. These iterations allow the core network server 302, base station 120, and / or UE 110 to dynamically modify the DNN used to process information exchanged via E2E communications based on changing operating conditions, as further described.

[0228] Fig.24 The signaling and control transaction diagram 2400 illustrates a fourth example of signaling and control transactions for E2E ML for a wireless network. In an implementation, aspects of the signaling and control transaction diagram 2400 are combined with Fig.21 Signaling and control transaction diagram 2100, Fig. 22 The signaling and control transaction diagram 2200 and / or Fig.23 The signaling and control transaction diagram 2300 operates as, and / or as a continuation thereof.

[0229] As shown, at 2405, UE 110 transmits a request for a machine learning configuration (ML configuration) to core network server 302 via base station 120 at 2410. For illustration, UE 110 transmits a request to base station 120, wherein the request indicates a specific ML configuration (such as a specific neural network formation configuration), and base station 120 relays or forwards the request to core network server 302. In response to detecting the invocation of the application, UE 110 sometimes determines performance requirements associated with the application and identifies an ML configuration designed to meet the performance requirements (e.g., by analyzing the reference Fig.12 UE 110 then requests ML configuration, such as by sending an indication of ML configuration in a NAS message for establishing E2E communication. In various implementations, the E2E communication corresponds to a QoS flow (e.g., QoS flow 1810, QoS flow 1812, QoS flow 1814). In some implementations, UE 110 includes a QoS identifier, such as a 5QI or a QCI, in the request.

[0230] At 2415, the core network server 302 determines the E2E ML configuration based on the ML configuration requested by the UE 110 through the E2E ML controller 318. For example, the E2E ML controller 318 analyzes any combination of ML capabilities, current operating environment, wireless network resource partitioning, performance requirements, etc. to determine the E2E ML configuration. In some implementations, the E2E ML controller partitions the E2E ML configuration based on the ML configuration requested by the UE 110, such that at least one partition in the E2E ML configuration corresponds to the ML configuration requested by the UE.

[0231] At 2420, core network server 302 communicates the E2E ML configuration, such as by communicating the first neural network formation configuration to the first device (e.g., base station 120), communicating the second neural network formation configuration to the second device (e.g., UE 110 via base station 120 at 2425), etc. In communicating the E2E ML configuration, core network server 302 sometimes indicates to UE 110 to use the ML configuration requested at 2405. Sometimes, core network server 302 identifies a third neural network configuration corresponding to forming a DNN at core network server 302. As an example, core network server 302 communicates a corresponding neural network formation configuration corresponding to a portion of the E2E ML configuration, such as by communicating a mapping to a reference UE 110. Figure 7 The index value in the entry of the neural network table.

[0232] At 2430, the core network server 302 forms a DNN (e.g., DNN 1912, DNN 1914, DNN 2020, DNN 2030) based on the E2E ML configuration determined at 2415. For example, referring to Fig.12 , the core network server accesses the neural network table to obtain parameters for forming the DNN. When forming the DNN, the core network server 302 forms a new DNN as part of establishing E2E communication and / or updates an existing DNN for exchanging information using E2E communication. In an implementation, the DNN formed by the core network server corresponds to a portion of an E2E ML configuration, as further described.

[0233] Similarly, at 2435, base station 120 forms a DNN (e.g., DNN 1910, DNN 1916, DNN 2018, DNN 2022, DNN 2028, DNN 2032) based on the E2E ML configuration determined at 2315. For example, referring to Fig.12 , base station 120 accesses the neural network table to obtain parameters for forming the DNN. In forming the DNN, the base station forms a new DNN as part of establishing E2E communication and / or updates an existing DNN for exchanging information using E2E communication. In an implementation, the DNN formed by base station 120 corresponds to a portion of an E2E ML configuration, as further described.

[0234] At 2440, UE 110 forms a DNN (e.g., DNN 1908, DNN 1918, DNN 2016, DNN 2024, DNN 2026, DNN 2034) based on the E2E ML configuration determined at 2415 and / or the ML configuration requested at 2405. For example, referring to Fig.12 , UE 110 accesses the neural network table to obtain parameters for forming the DNN. When forming the DNN, UE 110 forms a new DNN as part of establishing E2E communication and / or updates an existing DNN for exchanging information using E2E communication. In an implementation, the DNN formed by UE 110 corresponds to a portion of an E2E ML configuration, as further described.

[0235] Then, at 2445, at least the core network server 302, the base station 120, and the UE 110 use the DNN to process communications, wherein the communications correspond to information exchanged via the E2E communications. Figure 6 , the DNN at the device processes uplink and / or downlink information. Alternatively or additionally, the DNN at the device processes information exchanged via the QoS flow (e.g., Fig.18 ).

[0236] Having described signaling and control transactions that can be used for E2E ML for wireless networks, consider now some example methods that can be used for E2E ML for wireless networks in accordance with one or more implementations. Example Method

[0237] According to one or more aspects of the configuration of the neural network, reference Fig.25 and Fig.26 Example methods 2500 and 2600 are described. The order in which the method boxes are described is not intended to be construed as limiting, and any number of the method boxes may be skipped or combined in any order to implement a method or an alternative method. In general, any components, modules, methods, and operations described herein may be implemented using software, firmware, hardware (e.g., fixed logic circuits), manual processing, or any combination thereof. Some operations of the example methods may be described in the general context of executable instructions stored on a computer-readable storage memory locally and / or remotely from a computer processing system, and implementations may include software applications, programs, functions, and the like. Alternatively or additionally, any functions described herein may be performed at least in part by one or more hardware logic components, such as, but not limited to, a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), an application specific standard product (ASSP), a system on a chip (SoC), a complex programmable logic device (CPLD), and the like.

[0238] Fig.25 An example method 2500 for E2E ML for a wireless network is illustrated. In some implementations, the operations of the method 2500 are performed by an E2E ML controller, such as the E2E ML controller 318 implemented by the core network server 302.

[0239] At 2505, the E2E ML controller obtains capabilities of at least two devices utilized in the end-to-end communication. For example, the E2E ML controller implemented at a core network server (e.g., E2E ML controller 318, core network server 302) obtains ML capabilities from a base station (e.g., base station 120) and / or a UE (e.g., UE 110). The ML capabilities can include any combination of ML specific capabilities and / or processing capabilities, such as supported ML architectures, number of layers supported, available processing power, memory limitations, available power budget, fixed point processing versus floating point processing, maximum kernel size, computing power, etc. In some implementations, the E2E ML controller receives the capabilities in response to sending a request for capabilities from a device, such as at Fig. 22 Those described in 2205 of .

[0240] At 2510, the E2E ML controller determines an E2E ML configuration for processing information exchanged through the end-to-end communication based on the capabilities of at least two devices. For example, the E2E ML controller implemented at a core network server (e.g., E2E ML controller 318, core network server 302) analyzes ML capabilities from a base station (e.g., base station 120) and / or a UE (e.g., UE 110) to determine an E2E ML configuration, such as in Fig.21 2115 of Fig. 22 2230 of Fig.23 2310 of Fig.24 Those described at 2415 of . In some implementations, the E2E ML controller receives the metrics and analyzes current operating conditions to determine the E2E ML configuration.

[0241] Sometimes, the E2E ML controller receives the QoS profile and analyzes the QoS requirements, QoS parameters, and / or QoS characteristics, etc., to determine the E2E ML configuration. Alternatively or additionally, the E2E ML controller identifies one or more partitions as the E2E ML configuration to distribute functions and / or processing associated with the E2E ML configuration across one or more devices (e.g., core network server 302, base station 120, UE 110). For example, the E2E ML controller analyzes the neural network table and determines that a first neural network corresponding to a first portion of the E2E ML configuration forms a configuration, a second neural network corresponding to a second portion of the E2E ML configuration forms a configuration, and the like.

[0242] In some implementations, the E2E ML controller determines a distributed ML configuration to form a distributed DNN. For example, the E2E ML controller determines a first NN formation configuration corresponding to a first partition of the distributed DNN, a second NN formation configuration corresponding to a second partition of the distributed DNN, etc., wherein the corresponding device implements and / or forms the corresponding portion of the distributed DNN.

[0243] At 2515, the E2E ML controller directs at least two devices to process information exchanged through the end-to-end communication by forming a deep neural network based on the end-to-end machine learning configuration. For example, the E2E ML controller (e.g., E2E ML controller 318) transmits the corresponding neural network formation configuration to the corresponding device through the core network server (e.g., core network server 302), as shown in Fig.21 2125 of Fig. 22 2235 of Fig.23 2315 of Fig.24 Those described at 2420 of . In an implementation, the DNN corresponds to a distributed DNN.

[0244] Fig.26 An example method 2600 for E2E ML for a wireless network is illustrated. In some implementations, the operations of the method 2600 are performed by, for example, Figure 1 The UE 110 is used to execute.

[0245] At 2605, the UE transmits one or more capabilities supported by the UE. In one example, the UE (e.g., UE 110) transmits one or more ML capabilities to a core network server (e.g., core network server 302) via a base station (e.g., base station 120) and in response to receiving a request for ML capabilities, such as in Fig. 22 Those described at 2205 and 2220 of . Alternatively or additionally, the UE transmits processing capabilities as further described. In another example, the UE (e.g., UE 110) transmits one or more ML capabilities in response to detecting an invocation of an application. For example, in response to detecting an invocation of an application, UE 110 sends a request to establish an E2E connection for the application, and includes one or more ML capabilities in the request.

[0246] At 2610, the UE receives a neural network formation configuration based on the E2E ML configuration for processing information exchanged through the end-to-end communication. For example, the UE (e.g., UE 110) receives a neural network formation configuration corresponding to a portion of the E2E ML configuration from a core network server and through a base station (e.g., core network server 302, base station 120), such as in Fig.21 2125 of Fig. 22 2235 of Fig.23 2315 of Fig.24 Those described in 2420.

[0247] At 2615, the UE forms a deep neural network using the neural network formation configuration. For example, as in Fig.21 2145 Fig. 22 2255 of Fig.23 2335 and Fig.24 As described at 2440 of the present invention, the UE (e.g., UE 110) forms a DNN, such as by accessing a neural network table to obtain parameters and / or architecture configuration. In response to forming the DNN, at 2620, the UE uses the DNN to process information exchanged via E2E communication, such as in Fig.21 2150 of Fig. 22 2260 of Fig.23 2340 of Fig.24 Those described in 2445.

[0248] describe Fig.25and 26 The order of the method blocks is not intended to be construed as a limitation, and any number of the described method blocks can be combined in any order to implement a method, or an alternate method.

[0249] Generally, any components, modules, methods, and operations described herein may be implemented using software, firmware, hardware (e.g., fixed logic circuitry), manual processing, or any combination thereof. Some operations of the example methods may be described in the general context of executable instructions stored on a computer-readable storage memory local and / or remote to a computer processing system, and implementations may include software applications, programs, functions, and the like. Alternatively or additionally, any functions described herein may be performed, at least in part, by one or more hardware logic components, such as, but not limited to, a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), an application specific standard product (ASSP), a system on a chip (SoC), a complex programmable logic device (CPLD), and the like.

[0250] Although techniques and devices for E2E ML for wireless networks have been described in language specific to features and / or methods, it should be understood that the subject matter of the appended claims is not necessarily limited to the specific features or methods described. Rather, these specific features and methods are disclosed as example implementations of E2E ML for wireless networks.

[0251] In the following, several examples are described.

[0252] Example 1: A method performed by an end-to-end machine learning controller for determining an end-to-end machine learning configuration for processing information exchanged through end-to-end communication in a wireless network, the method comprising: obtaining, by the end-to-end machine learning controller, capabilities of at least two devices utilized in the end-to-end communication; determining, based on the capabilities of the at least two devices, the end-to-end machine learning configuration for processing information exchanged through the end-to-end communication; and guiding the at least two devices to process the information exchanged through the end-to-end communication by forming a deep neural network based on the end-to-end machine learning configuration.

[0253] Example 2: A method according to Example 1, wherein determining the end-to-end machine learning configuration further includes: dividing the end-to-end machine learning configuration across the at least two devices by: determining a first neural network formation configuration corresponding to a first part of the end-to-end machine learning configuration based on capabilities of a first device among the at least two devices; and determining a second neural network formation configuration corresponding to a second part of the end-to-end machine learning configuration based on capabilities of a second device among the at least two devices, wherein the capabilities of the first device include available processing power of the first device, and wherein the capabilities of the second device include available processing power of the second device.

[0254] Example 3: A method according to Example 1 or Example 2, wherein the end-to-end machine learning configuration is a first end-to-end machine learning configuration, and the method further includes: obtaining one or more metrics indicating a current operating environment for the end-to-end communication; identifying a second end-to-end machine learning configuration based at least on the one or more metrics indicating the current operating environment; and guiding the at least two devices to update the one or more deep neural networks based on the second end-to-end machine learning configuration.

[0255] Example 4: A method according to any one of Examples 1 to 3, wherein determining the end-to-end machine learning configuration includes: determining a first neural network formation configuration for a deep neural network on a user device side as a first part of the end-to-end machine learning configuration; determining a second neural network formation configuration for a deep neural network on a base station side as a second part of the end-to-end machine learning configuration; and determining a third neural network formation configuration for a deep neural network on a core network server side as a third part of the end-to-end machine learning configuration.

[0256] Example 5: A method according to any one of Examples 1 to 4, wherein determining the end-to-end machine learning configuration further includes: obtaining at least one quality of service parameter or quality of service characteristic associated with the end-to-end communication; and determining the end-to-end machine learning configuration based at least in part on the at least one quality of service parameter or quality of service characteristic.

[0257] Example 6: A method according to Example 5, wherein the at least one quality of service parameter or quality of service characteristic includes at least one of the following: priority level; packet delay budget; packet error rate; maximum data burst size; or averaging window.

[0258] Example 7: A method according to any one of Examples 1 to 6, wherein the end-to-end machine learning configuration is a first end-to-end machine learning configuration, the end-to-end communication is a first quality of service flow, the deep neural network is a first deep neural network, and the method further includes: determining to establish a second quality of service flow between the at least two devices; determining a second end-to-end machine learning configuration for processing information exchanged through the second quality of service flow; and guiding the one or more devices to process the information exchanged through the second quality of service flow by forming a second deep neural network based on the second end-to-end machine learning configuration.

[0259] Example 8: A method according to any one of Examples 1 to 7, wherein the end-to-end communication is associated with exchanging information, and the exchanging information is associated with one of the following: augmented reality applications; virtual reality applications; audio streaming applications; real-time gaming applications; video streaming applications; Internet Protocol voice applications; social media applications; file transfer; vehicle-to-everything communications; Internet of Things communications; automation; or remote control.

[0260] Example 9: A method according to any one of Examples 1 to 8, wherein the at least two devices include a user device, and wherein determining the end-to-end machine learning configuration further includes: receiving a request for a specific neural network formation configuration from the user device; and determining the end-to-end machine learning configuration based on the specific neural network formation configuration.

[0261] Example 10: A method according to any one of Examples 1 to 9, wherein obtaining the capabilities of the one or more devices includes obtaining machine learning capabilities of the at least two devices.

[0262] Example 11: A method according to any one of Examples 1 to 10, wherein determining the end-to-end machine learning configuration includes: determining an architecture configuration and one or more parameter configurations that define a deep neural network as the end-to-end machine learning configuration.

[0263] Example 12: A method according to any one of Examples 1 to 10, wherein determining the end-to-end machine learning configuration includes: analyzing one or more metrics of a current operating environment; and determining one or more parameter configurations that define updates to a deep neural network as the end-to-end machine learning configuration based in part on the one or more metrics.

[0264] Example 13: A method performed by a user device, the method comprising: transmitting, by the user device, one or more capabilities supported by the user device; receiving a neural network formation configuration based on an end-to-end machine learning configuration for processing information exchanged through the end-to-end communication; using the neural network formation configuration to form a deep neural network; and using the deep neural network to process information exchanged through the end-to-end communication.

[0265] Example 14: The method of Example 13, wherein transmitting one or more capabilities supported by the user device comprises transmitting at least one of: a maximum kernel size capability; a memory limit; or a computing capability.

[0266] Example 15: A method according to Example 13 or Example 14, wherein the end-to-end communication is a first quality of service flow associated with interactive communication, the neural network formation configuration is a first neural network formation configuration, the deep neural network is a first deep neural network, the information exchanged through the end-to-end communication includes information associated with the interactive communication, and the method further includes: transmitting a request to establish a second quality of service flow, receiving a second neural network formation configuration for processing information exchanged through the second quality of service flow; forming a second deep neural network based on the second neural network formation configuration; and using the second deep neural network to process information exchanged through the second quality of service flow.

[0267] Example 16: A network entity comprising: a processor; and a computer-readable storage medium comprising instructions, wherein the instructions implement an end-to-end machine learning controller for performing any of the methods described in Examples 1 to 12.

[0268] Example 17: A user device comprising: a processor; and a computer-readable storage medium comprising instructions, wherein the instructions, in response to being executed by the processor, are used to direct the user device to perform one of the methods described in Examples 13 to 15.

Claims

1. A method performed by an end-to-end machine learning controller for determining an end-to-end machine learning configuration for processing information exchanged via end-to-end communications in a wireless network, the method comprising: Obtaining machine learning capabilities of at least two devices involved in the peer-to-peer communication, the machine learning capability of a first device comprising a first available processing power of the first device, and the machine learning capability of a second device comprising a second available processing power of the second device; determining, based on the machine learning capabilities of the at least two devices, the end-to-end machine learning configuration for processing information exchanged through the end-to-end communication; as well as The at least two devices are directed to process the information exchanged through the end-to-end communication by forming a deep neural network based on the end-to-end machine learning configuration.

2. The method according to claim 1, wherein: Determining the end-to-end machine learning configuration further includes: Partitioning the end-to-end machine learning configuration across the at least two devices by: determining a first neural network formation configuration corresponding to a first portion of the end-to-end machine learning configuration based on a machine learning capability of the first device of the at least two devices; and A second neural network formation configuration corresponding to a second portion of the end-to-end machine learning configuration is determined based on a machine learning capability of the second device of the at least two devices.

3. The method according to claim 1, wherein: The end-to-end machine learning configuration is a first end-to-end machine learning configuration, and the method further comprises: obtaining one or more metrics indicative of a current operating environment for the peer-to-peer communication; identifying a second end-to-end machine learning configuration based at least on the one or more metrics indicative of the current operating environment; and The at least two devices are directed to update the one or more deep neural networks based on the second end-to-end machine learning configuration.

4. The method according to claim 1, wherein: Determining the end-to-end machine learning configuration includes: Determining a first neural network formation configuration for a user device-side deep neural network as a first part of the end-to-end machine learning configuration; determining a second neural network formation configuration for a base station-side deep neural network as a second part of the end-to-end machine learning configuration; and A third neural network formation configuration for a core network server-side deep neural network is determined as a third part of the end-to-end machine learning configuration.

5. The method according to claim 1, wherein: Determining the end-to-end machine learning configuration further includes: obtaining at least one quality of service parameter or quality of service characteristic associated with the end-to-end communication; and The end-to-end machine learning configuration is determined based at least in part on the at least one quality of service parameter or quality of service characteristic.

6. The method according to claim 5, wherein: The at least one quality of service parameter or quality of service characteristic comprises at least one of the following: Priority level; Packet delay budget; Packet error rate; Maximum data burst size; and Averaging window.

7. The method according to claim 1, wherein: The end-to-end machine learning configuration is a first end-to-end machine learning configuration, the end-to-end communication is a first quality of service flow, the deep neural network is a first deep neural network, and the method further comprises: Determining to establish a second quality of service flow between the at least two devices; determining a second end-to-end machine learning configuration for processing information exchanged via the second quality of service flow; and The at least two devices are directed to process the information exchanged through the second quality of service flow by forming a second deep neural network based on the second end-to-end machine learning configuration.

8. The method according to claim 1, wherein: The end-to-end communication is associated with exchange information, and the exchange information is associated with one of: Augmented reality applications; Virtual reality applications; Audio streaming applications; Real-time gaming applications; Video streaming applications; Voice over Internet Protocol applications; Social media applications; File transfer; Car-to-everything communication; IoT communications; automation; as well as Remote control.

9. The method according to claim 1, wherein: The at least two devices include a user device, and Wherein, determining the end-to-end machine learning configuration further includes: receiving a request from the user device for a particular neural network formation configuration; and The end-to-end machine learning configuration is determined based on the particular neural network formation configuration.

10. The method according to any one of claims 1 to 9, wherein: Determining the end-to-end machine learning configuration includes: Determining an architecture configuration and one or more parameter configurations defining the deep neural network as the end-to-end machine learning configuration; or Analyzing one or more metrics of a current operating environment, and determining, based in part on the one or more metrics, one or more parameter configurations defining updates to the deep neural network as the end-to-end machine learning configuration.

11. A method performed by a user equipment, the method comprising: transmitting one or more machine learning capabilities supported by the user device, the one or more machine learning capabilities supported by the user device comprising at least one of: Maximum kernel size capability, Memory limitations, and Computational capacity; receiving a neural network forming configuration based on an end-to-end machine learning configuration for processing information exchanged via the end-to-end communication; forming a deep neural network using the neural network forming configuration; as well as The information exchanged through the end-to-end communication is processed using the deep neural network.

12. The method according to claim 11, wherein: The end-to-end communication is a first quality of service flow associated with interactive communication, the neural network formation configuration is a first neural network formation configuration, the deep neural network is a first deep neural network, the information exchanged via the end-to-end communication includes information associated with the interactive communication, and the method further includes: transmitting a request to establish a second quality of service flow; receiving a second neural network forming configuration for processing information exchanged via the second quality of service flow; forming a second deep neural network based on the second neural network forming configuration; and The second deep neural network is used to process the information exchanged via the second quality of service flow.

13. A network entity, comprising: processor; as well as A computer-readable storage medium comprising instructions implementing an end-to-end machine learning controller for performing the method according to any one of claims 1 to 10.

14. A user equipment, comprising: processor; as well as A computer-readable storage medium comprising instructions, the instructions being responsive to being executed by the processor for directing the user equipment to perform the method according to claim 11 or claim 12.

Images