End-to-end deep neural network adaptation for edge computing

Abstract

Techniques and apparatuses are described for adaptive end-to-end, E2E, machine learning, ML, configuration for processing communications communicated over E2E communications. A network entity directs a user equipment (UE) and a base station participating in an E2E communication to implement the E2E communication by forming at least a portion of an E2E deep neural network, DNN, based on a first E2E ML configuration (805). The network entity determines to update the first E2E ML configuration based on a change in a participation mode of an edge computing server, ECS, in the E2E communication (815). The network entity identifies a second E2E ML configuration based on the change in the participation mode (820), and instructs the UE or the base station to update the portion of the E2E DNN using the second E2E ML configuration (825).

Description

Background Technology

[0001] The development of wireless communication systems is often driven by the need for higher data throughput and lower latency. As an example, the demand for data increases as more devices gain access to wireless communication systems. These evolving devices also perform data-intensive and / or compute-intensive applications that utilize and / or process more data than previous applications, such as streaming video applications with higher refresh rates, social media applications with higher resolutions, multiplayer gaming applications, and audio services with higher fidelity. To accommodate increased data usage and provide more computing power, evolving wireless communication systems incorporate additional computing devices and / or data storage resources.

[0002] As an example, mobile edge computing corresponds to computing applications, capabilities, and / or services deployed at the edge of a wireless network, such as the coverage area of ​​a fifth-generation (5G) and / or sixth-generation (6G) wireless network. Compared to cloud-based services accessed via the internet, mobile edge computing provides computing resources local to the base station and subsequently to the user equipment (UE) communicating with the base station. The locality of edge computing improves network response time by reducing data transmission latency.

[0003] Wireless networks sometimes migrate UEs between edge computing servers (ECS) and cloud-based services for application processing. For illustration, consider a UE operating within a wireless network. When a UE moves to a first coverage area with an available ECS, the wireless network can determine to utilize the computing resources provided by the ECS for the UE's application processing. However, as the UE moves away from the coverage area and / or ECS, the wireless network must redirect the UE's application processing to a different ECS and / or cloud-based service. Therefore, edge computing adds complexity as the wireless network dynamically redirects application processing between edge computing and cloud-based services. Summary of the Invention

[0004] This document describes techniques and apparatus for end-to-end (E2E) deep neural network (DNN) adaptation for edge computing. Various aspects describe the adaptive formation of an end-to-end E2E machine learning (ML) configuration for processing E2E deep neural networks (DNNs) transmitted via E2E communication. A network entity guides user equipment (UEs) and base stations participating in E2E communication to implement E2E communication by forming at least a portion of the E2E DNN based on a first E2E ML configuration. The network entity determines to update the first E2E ML configuration based on changes in the participation mode of edge computing servers (ECS) in the E2E communication. The network entity identifies a second E2E ML configuration based on the change in participation mode and guides the UE or base station to update a portion of the E2E DNN using the second E2E ML configuration.

[0005] In all aspects, the UE adaptively configures its E2E ML to process communications transmitted via E2E communication. The UE forms a deep neural network (DNN) using at least a first portion of a first E2E ML configuration for implementing the E2E DNN. The UE receives an indication that it updates the DNN using at least a second portion of a second E2E ML configuration based on a change in the participation mode of an edge computing server (ECS) in the E2E communication. The UE then updates the DNN using at least a second portion of the E2E ML configuration based on the ECS participation mode and uses the updated DNN to implement at least a portion of the E2E communication.

[0006] Details of one or more embodiments of an E2E DNN adaptive approach for edge computing are set forth in the accompanying drawings and the following description. Other features and advantages will be apparent from the specification and drawings, as well as from the claims. This summary is provided to introduce the subject matter further described in the detailed description and drawings. Therefore, this summary should not be construed as describing essential features, nor should it be used to limit the scope of the claimed subject matter. Attached Figure Description

[0007] The following describes in detail one or more aspects of the adaptive end-to-end (E2E) deep neural network (DNN) for edge computing. The same reference numerals are used to denote similar elements in different instances of the specification and figures:

[0008] Figure 1 The illustration shows an example environment that enables various aspects of E2E DNN adaptation for edge computing;

[0009] Figure 2 The illustration shows an example device diagram of a device capable of implementing various aspects of E2E DNN adaptation for edge computing;

[0010] Figure 3 The illustration shows example device diagrams of other devices capable of implementing various aspects of E2E DNN adaptation for edge computing;

[0011] Figure 4 The illustration shows an example operating environment in a wireless communication system utilizing multiple deep neural networks, based on various aspects of E2E DNN adaptation for edge computing.

[0012] Figure 5 The illustration shows an example of a configuration that generates multiple neural networks based on various aspects of an E2E DNN adapted for edge computing.

[0013] Figure 6The illustration shows an example operating environment that enables E2E DNN adaptation for edge computing based on various aspects;

[0014] Figure 7 The illustration shows another example operating environment that enables E2E DNN adaptation for edge computing based on various aspects;

[0015] Figure 8 The diagram illustrates an example transaction graph between various network entities that implement E2E DNN adaptation for edge computing;

[0016] Figure 9 The diagram illustrates other example transaction graphs between various network entities that implement E2E DNN adaptation for edge computing;

[0017] Figure 10 The illustration shows an example method for E2E DNN adaptation for edge computing; and

[0018] Figure 11 Another example approach for E2E DNN adaptation for edge computing is illustrated. Detailed Implementation

[0019] Edge computing, sometimes called mobile edge computing (MEC), provides wireless networks with the ability to improve communication with user equipment (UE) by using edge computing servers (ECS) to provide additional computing resources and / or reduce data transmission latency. As an example, a base station in a wireless network can connect to an ECS and provide edge computing services to the UE to improve data transmission latency. To illustrate, the close proximity of the ECS to the target device, relative to the proximity of a remote service or data center, helps improve the responsiveness of the wireless network by reducing data transmission latency with the UE. Reduced latency helps improve the performance and responsiveness of various applications accessing resources provided by edge computing, such as gaming applications, augmented reality (AR) applications, virtual reality (VR) applications, vehicle applications (e.g., road information, weather applications), real-time drone detection, and data analytics.

[0020] Device mobility adds complexity to managing edge computing resources. To illustrate, when a UE moves to a first coverage area connected to a first base station of an ECS, the wireless network can incorporate edge computing for processing the UE's application data. When the UE moves out of the first coverage area and into a second coverage area without a second base station of an ECS, the wireless network must redirect the application data to a cloud-based service.

[0021] The DNN provides solutions for complex processing, such as functionality associated with changing participation in edge computing in E2E communication (e.g., adding, omitting, and aggregating with cloud-based services) supporting data sessions with a wireless network. Various aspects train the DNN to process communications delivered via E2E communication based on the participation patterns of edge computing in E2E communication. Then, as the participation patterns of edge computing in E2E communication change, the wireless network can guide the devices participating in the E2E communication to form or update the E2E DNN. In some aspects, the wireless network dynamically reconfigures the E2E DNN by modifying various parameter configurations (e.g., coefficients, kernel size, weights) and / or various architectural configurations and / or layer computation patterns (e.g., adding convolutional layers, reducing the number of convolutional layers, increasing or decreasing the downsampling of data performed by layers, reducing the number of fully connected layers, increasing the number of fully connected layers), to improve processing resolution or reduce processing computation time. This provides wireless networks with the flexibility to dynamically adapt the E2E DNN when E2E communication is directed to or from an edge computing server (ECS), such as when the UE moves into a coverage area that includes an ECS or when the UE moves out of the coverage area. It also allows wireless networks to reconfigure the E2E DNN using an edge computing-optimized architecture when E2E communication includes an ECS, and to reconfigure the E2E DNN using a cloud-based computing-optimized architecture when E2E communication omits an ECS, thereby improving processing resolution, processing time, latency, and other aspects based on ECS participation patterns.

[0022] This document describes aspects of E2E DNN adaptation for edge computing, which allows the system to process communications and dynamically reconfigure the DNN used in E2E communications as an endpoint for E2E communication changes. Each aspect describes the formation and / or adaptation of an end-to-end E2E machine learning (ML) configuration that forms an E2E deep neural network (DNN) for processing communications transmitted over E2E communications. A network entity guides user equipment (UE) and base stations participating in E2E communications to implement E2E communications by forming at least a portion of the E2E DNN based on a first E2E ML configuration. The network entity determines to update the first E2E ML configuration based on changes in the participation mode of the ECS in the E2E communications. The network entity identifies a second E2E ML configuration based on the changes in the participation mode and guides the UE or base station to use the second E2E ML configuration to update a portion of the E2E DNN.

[0023] In all aspects, the UE adaptively configures its E2E ML to process communications transmitted via E2E communication. The UE uses at least a first portion of a first E2E ML configuration for implementing the E2E DNN to form a deep neural network (DNN). The UE receives an indication that it updates the DNN using at least a second portion of a second E2E ML configuration based on a change in the ECS participation mode during E2E communication. The UE then updates the DNN using at least the second portion of the E2E ML configuration based on the ECS participation mode and uses the updated DNN to implement at least a portion of the E2E communication.

[0024] Example Environment

[0025] Figure 1 The illustration includes an example environment 100 for User Equipment 110 (UE 110), which is capable of communicating with Base Station 120 (illustrated as Base Station 121 and 122) via one or more wireless communication links 130 (wireless links 130) (illustrated as wireless links 131 and 132). For simplicity, UE 110 is implemented as a smartphone, but can be implemented as any suitable computing or electronic device, such as a mobile communication device, modem, cellular phone, gaming device, navigation device, media device, laptop, desktop computer, tablet, smart device, vehicle-based communication system, or Internet of Things (IoT) device, such as a sensor or actuator. Base Station 120 can be implemented in macrocells, microcells, small cells, picocells, distributed base stations (one or more), or any combination thereof (e.g., 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.).

[0026] Base station 120 communicates with user equipment 110 using wireless links 131 and 132, which can be implemented as any suitable type of wireless link. Wireless links 131 and 132 include control and data communications, such as downlinks transmitting data and control information from base station 120 to user equipment 110, uplinks transmitting other data and control information from user equipment 110 to base station 120, or both. Wireless link 130 may include one or more wireless 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 3GPP LTE, 5G NR, etc. Multiple wireless links 130 can be aggregated in carrier aggregation or multi-connectivity technologies to provide higher data rates for UE 110. Multiple wireless links 130 from multiple base stations 120 can be configured for Co-op Multipoint communication with UE 110.

[0027] Base stations 120 collectively form a radio access network 140 (e.g., RAN, Evolved Universal Terrestrial Radio Access Network, E-UTRAN, 5G NR RAN, or NR RAN). Base stations 121 and 122 in RAN 140 are connected to the core network 150. Base stations 121 and 122 are connected to the core network 150 at locations 102 and 104, respectively, via the NG2 interface for control plane signaling and the NG3 interface for user plane data communication when connected to the 5G core network, or via the S1 interface for both control plane signaling and user plane data communication when connected to the Evolved Packet Core (EPC) network. Base stations 121 and 122 are capable of communicating using the Xn Application Protocol (XnAP) via the Xn interface or the X2 Application Protocol (X2AP) via the X2 interface to exchange user plane data and control plane information at location 106. User equipment 110 can connect to a public network, such as the Internet 160, via the core network 150 to interact with remote services 170. Remote services 170 refers to computing, communication, and storage devices used to provide any of a variety of services, including interactive voice or video communication, file transfer, streaming voice or video, and other technical services implemented in any way, such as voice calls, video calls, website access, messaging services (e.g., text messaging or multimedia messaging), photo file transfer, enterprise software applications, social media applications, video games, streaming video or audio services, and podcasts.

[0028] RAN 140 also includes one or more edge computing servers 180 (ECS 180), illustrated here as edge computing server 181, and edge computing server 182 (ECS 181, ECS 182), which provide edge computing resources for processing application data of UE 110. ECS 181 uses an Xe interface, shown at interface 191, connected to base station 121, and connected to core network 150 via core network interface 192. Similarly, ECS 182 uses an Xe interface, shown at interface 193, connected to base station 122, and connected to core network 150 via core network interface 194. In various aspects, core network 150 manages and / or grants access to resources of ECS 181 and / or ECS 182. Alternatively or additionally, core network 150 manages the mobility of applications between ECS 181, ECS 182, and / or remote service 170, as well as the associated data and context of the applications. As an example, when UE 110 switches from a serving cell base station (e.g., base station 121) to a neighboring base station (e.g., base station 122), the core network 150 server passes the application from ECS 181, along with any data and context associated with the application, to ECS 182. Therefore, in various aspects, the core network 150 includes ECS management functions.

[0029] Example device

[0030] Figure 2 Figure 200 illustrates an example device, one of a UE 110 and a base station 120, capable of implementing various aspects of E2E DNN adaptation for edge computing. Figure 3 Example device diagram 300 illustrates a core network server 302 and an ECS 180 (e.g., ECS 181, ECS 182) capable of implementing various aspects of E2E DNN adaptation for edge computing. The UE 110, base station 120, core network server 302, and / or ECS 180 may include, for clarity, [details omitted]. Figure 2 or Figure 3 Other functions and interfaces omitted.

[0031] UE 110 includes an antenna 202, a radio frequency front-end 204 (RF front-end 204), and a radio transceiver (e.g., LTE transceiver 206 and / or 5G NR transceiver 208) for communicating with a base station 120 in RAN 140. The RF front-end 204 of UE 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 communication. The antenna 202 of UE 110 may include an array of multiple antennas configured similarly or differently from each other. The antenna 202 and RF front-end 204 can be tuned to 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 / or the 5G NR transceiver 208. Additionally, antenna 202, RF front-end 204, LTE transceiver 206, and / or 5G NR transceiver 208 can 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 can be implemented for operation in sub-gigahertz, sub-6GHz, and / or higher frequency bands as defined by the 3GPP LTE and 5G NR communication standards.

[0032] User equipment 110 also includes one or more processors 210 and a computer-readable storage medium 212 (CRM 212). Processor 210 may be a single-core or multi-core processor made of various materials such as silicon, polysilicon, high-k dielectrics, copper, etc. The computer-readable storage medium described herein excludes propagated signals. CRM 212 may include any suitable memory or storage device that can be used to store device data 214 of UE 110, such as random access memory (RAM), static RAM (SRAM), dynamic RAM (DRAM), non-volatile RAM (NVRAM), read-only memory (ROM), or flash memory. Device data 214 includes user data, multimedia data, beamforming codebooks, applications, neural network (NN) tables, and / or the operating system of UE 110 executable by processor (one or more) 210 to enable user plane communications, control plane signaling, and user interaction with user equipment 110.

[0033] In various aspects, CRM 212 includes a neural network table 216 that stores various architectures and / or parameter configurations for forming neural networks, such as, by example rather than limitation, specifying fully connected layer neural network architectures, convolutional layer neural network architectures, recurrent neural network layers, multi-connected hidden neural network layers, input layer architectures, output layer architectures, multiple nodes utilized by the neural network, coefficients (e.g., weights and biases) utilized by the neural network, kernel parameters, multiple filters utilized by the neural network, stride / pooling configurations utilized by the neural network, activation functions for each neural network layer, interconnections between neural network layers, neural network layers to be skipped, etc. Therefore, neural network table 216 includes arbitrary combinations of neural network forming configuration elements (NN forming configuration elements) such as architectures and / or parameter configurations, which can be used to create neural network forming configurations (NN forming configurations) comprising combinations of one or more NN forming configuration elements that define and form a DNN. In some aspects, a single index value in neural network table 216 maps to a single NN forming configuration element (e.g., a 1:1 correspondence). Alternatively or additionally, individual index values ​​of neural network table 216 are mapped to NN forming configurations (e.g., combinations of NN forming configuration elements). In some embodiments, the neural network table includes input characteristics for each NN forming configuration element and / or NN forming configuration, wherein the input characteristics describe properties related to the training data used to generate the NN forming configuration elements and / or NN forming configurations, as further described.

[0034] 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 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 via index values, and forms the DNN using NN formation configuration elements specified by the NN formation configuration. This includes updating the DNN using any combination of architectural changes and / or parameter changes to the DNN as further described, such as minor changes to the DNN involving updating parameters and / or major changes reconfiguring the nodes and / or layer connections of the DNN. In implementations, the UE Neural Network Manager forms multiple DNNs to handle wireless communications (e.g., downlink communications, uplink communications).

[0035] Figure 2The illustrated device diagram of base station 120 includes a single network node (e.g., gNode B). The functionality of base station 120 can be distributed across multiple network nodes or devices, and can be distributed in any manner suitable for performing the functions described herein. Base station 120 includes an antenna 252, a radio frequency front-end 254 (RF front-end 254), and one or more radio transceivers (e.g., one or more LTE transceivers 256 and / or one or more 5G NR transceivers 258) for communicating with UE 110. The RF front-end 254 of base station 120 is capable of coupling or connecting the LTE transceivers 256 and 5G NR transceivers 258 to antenna 252 to facilitate various types of wireless communication. Antenna 252 of base station 120 may include an array of multiple antennas configured in similar or different ways. Antenna 252 and RF front-end 254 can be tuned to and / or tunable to one or more frequency bands defined by the 3GPP LTE and 5G NR communication standards and implemented by the LTE transceivers 256 and / or 5G NR transceivers 258. Additionally, antenna 252, RF front end 254, LTE transceiver 256 and / or 5G NR transceiver 258 can be configured to support beamforming such as Massive-MIMO for transmitting and receiving communications with UE 110.

[0036] Base station 120 also includes one or more processors 260 and computer-readable storage medium 262 (CRM 262). Processor 260 may be a single-core or multi-core processor composed of various materials such as silicon, polysilicon, high-k dielectric, copper, etc. CRM 262 may include any suitable memory or storage device that can be used to store device data 264 of base station 120, such as random access memory (RAM), static RAM (SRAM), dynamic RAM (DRAM), non-volatile RAM (NVRAM), read-only memory (ROM), or flash memory. Device data 264 includes network scheduling data, radio resource management data, beamforming codebooks, applications, and / or the operating system of base station 120, which can be executed by one or more processors 260 to enable communication with UE 110.

[0037] CRM 262 also includes a base station manager 266. Alternatively or additionally, the base station manager 266 may be implemented, in whole or in part, as hardware logic or circuitry 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 to communicate with the UE 110 and with a core network such as the core network 150.

[0038] 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 may be implemented, in whole or in part, as hardware logic or circuitry integrated or separate from other components of the base station 120. In at least some aspects, such as by selecting a combination of NN formation configuration elements to form a DNN for processing edge computing communications and / or cloud-based communications, the BS Neural Network Manager 268 selects an NN formation configuration 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 an NN formation configuration based on that feedback. Alternatively or additionally, the BS Neural Network Manager 268 receives neural network formation configuration guidance from elements of the core network 150 via the core network interface 276 or the inter-base station interface 274 and forwards the NN formation configuration guidance to the UE 110. In some aspects, the BS Neural Network Manager 268 selects an NN formation configuration in response to determining whether to add edge computing to or remove edge computing from E2E communications.

[0039] CRM 262 includes a training module 270 and a neural network table 272. In an implementation, base station 120 manages the NN formation configuration and deploys it to UE 110. Alternatively or additionally, base station 120 maintains the neural network table 272. Training module 270 uses known input data to teach and / or train the DNN. For example, training module 270 trains one or more DNNs for various purposes, such as processing communications transmitted through wireless communication systems (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, processing edge computing communications, processing cloud-based computing communications, and aggregating edge computing communications with cloud-based computing communications). This includes training one or more DNNs 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).

[0040] In this implementation, the training module 270 extracts the learned parameter configuration 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 to the neural network table 272. The extracted parameter configuration includes any combination of information defining the behavior of the neural network, such as node connections, coefficients, active layers, weights, biases, pooling, etc.

[0041] The neural network table 272 stores multiple different NN formation configuration elements and / or NN formation configurations generated using the training module 270. In some embodiments, the neural network table includes input characteristics for each NN formation configuration element and / or NN formation configuration, wherein the input characteristics describe attributes related to the training data used to generate the NN formation configuration elements and / or NN formation configurations. For example, input characteristics include, but are not limited to, edge computing participation modes, endpoint participation modes, cloud-based computing participation modes (e.g., addition, omission, aggregation), 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 transmit error rate, second transmit error rate, maximum retransmission), latency, radio link control (RLC), automatic repeat request (ARQ) metrics, received signal strength (RSS), uplink SINR, timing measurements, error metrics, UE capabilities, BS capabilities, power patterns, Internet Protocol (IP) layer throughput, end-to-end latency, end-to-end packet loss rate, etc. Therefore, input characteristics sometimes include layer 1, layer 2, and / or layer 3 metrics. In some implementations, a single index value of neural network table 272 is mapped to a single NN to form a configuration element (e.g., a 1:1 correspondence). Alternatively or additionally, individual index values ​​of neural network table 272 are mapped to NN forming configurations (e.g., combinations of NN forming configuration elements).

[0042] In one implementation, base station 120 synchronizes neural network table 272 with neural network table 216 such that the NN forming configuration elements and / or input characteristics stored in one neural network table are copied in the second neural network table. Alternatively or additionally, base station 120 synchronizes neural network table 272 with neural network table 216 such that the NN forming configuration elements and / or input characteristics stored in one neural network table represent complementary functions in the second neural network table (e.g., NN forming configuration elements for transmitter path processing in the first neural network table, and NN forming configuration elements for receiver path processing in the second neural network table).

[0043] Base station 120 also includes an inter-base station interface 274, such as an Xn and / or X2 interface, which base station manager 266 configures to exchange user plane data, control plane information, and other data / information with other base stations to manage communication between base station 120 and UE 110. Base station 120 includes a core network interface 276, which base station manager 266 configures to exchange user plane data, control plane information, and / or other data / information with core network functions and / or entities.

[0044] exist Figure 3In the core network 150, core network server 302 can provide all or part of the functions, entities, services, and / or gateways. Each function, entity, service, and / or gateway in core network 150 can be provided as a service distributed across multiple servers or embodied on a dedicated server within core network 150. For example, core network server 302 can provide all or part of the services or functions of 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. Core network server 302 is illustrated as being embodied on a single server including one or more processors 304 and computer-readable storage medium 306 (CRM306). Processor 304 can be a single-core processor or a multi-core processor composed of various materials such as silicon, polysilicon, high-k dielectric, copper, etc. CRM 306 may include any suitable memory or storage device that can be used to store device data 308 of the core network server 302, such as random access memory (RAM), static RAM (SRAM), dynamic RAM (DRAM), non-volatile RAM (NVRAM), read-only memory (ROM), hard disk drive, or flash memory. Device data 308 includes data that can be executed by one or more processors 304 to support core network functions or entities, and / or the operating system of the core network server 302.

[0045] CRM 306 also includes one or more core network applications 310, which in one implementation are embodied on CRM 306 (as shown). The one or more core network applications 310 may implement functions such as UPF, AMF, S-GW, P-GW, MME, ePDG, ECS management, etc. Alternatively or additionally, the one or more core network applications 310 may be implemented, in whole or in part, as hardware logic or circuitry integrated or separate from other components of the core network server 302.

[0046] CRM 306 also includes a core network neural network manager 312, which manages the formation of NN configurations for processing communications transmitted between UE 110 and base station 120, such as adding or omitting E2E communications of ECS for application processing, and the formation configuration of the DNN. In various aspects, the core network neural network manager 312 analyzes various E2E communication endpoint connection configurations (e.g., including / excluding ECS ​​180, including / excluding remote service 170, aggregation of ECS 180 communications and remote service 170 communications), and selects an end-to-end machine learning configuration (E2E ML configuration) that can be used to form the end-to-end deep neural network (E2E DNN) for processing communications in E2E communications based on the added and / or omitted endpoints. In various aspects, the core network neural network manager 312 selects one or more NN formation configurations within neural network table 316 to indicate the determined E2E ML configuration.

[0047] In some implementations, the core network neural network manager 312 analyzes various criteria, such as current signal channel conditions (e.g., reported by base station 120, other radio access points, or UE 110 (via base station or other radio access points)), the capabilities of base station 120 (e.g., antenna configuration, cell configuration, MIMO capability, radio capability, processing capability), and the capabilities of UE 110 (e.g., antenna configuration, MIMO capability, radio capability, processing capability). For example, base station 120 obtains various criteria and / or link quality indications during communication with the UE and forwards these criteria and / or link quality indications to the core network neural network manager 312. The core network neural network manager selects an E2E ML configuration based on these criteria and / or indications to improve the accuracy of the DNN processing the communication (e.g., lower bit error rate, higher signal quality). The core network neural network manager 312 then transmits the E2E ML configuration to base station 120 and / or UE 110, such as by transmitting the index of the neural network table. In this implementation, the core network neural network manager 312 receives feedback from the base station 120 from the UE and / or BS, and selects an updated E2E ML configuration based on the feedback.

[0048] CRM 306 includes a training module 314 and a neural network table 316. In one implementation, a core network server 302 manages E2E ML configurations and / or portions of partitionable E2E ML configurations and deploys them to multiple devices (e.g., UE 110, base station 120) in a wireless communication system. Alternatively or additionally, the core network server maintains the neural network table 316 externally to CRM 306. The training module 314 teaches and / or trains a DNN using known input data. For example, the training module 314 trains one or more DNNs to handle different types of pilot communications transmitted through the wireless communication system. This includes offline and / or online training of one or more DNNs. In one implementation, the training module 314 extracts learned NN forming configurations and / or learned NN forming configuration elements from the DNN and stores the learned NN forming configuration elements in the neural network table 316, such as NN forming configurations that can be selected by the core network neural network manager 312 as E2E ML configurations to form an E2E DNN, as further described. 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 is mapped 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 is mapped to an NN formation configuration (e.g., a combination of NN formation configuration elements).

[0049] In some implementations, the training module 314 of the core network neural network manager 312 generates complementary NN formation configurations and / or NN formation configuration elements to those stored in the neural network table 216 at UE 110 and / or the neural network table 272 at base station 121. As an example, the training module 314 uses the NN formation configurations and / or NN formation configuration elements to generate the neural network table 316, which exhibit significant variations in architecture and / or parameter configuration compared to the moderate and / or low variations used to generate the neural network table 272 and / or 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 high-precision neural networks at the cost of increased processing complexity and / or time.

[0050] Neural network table 316 stores multiple different NN formation configuration elements generated using 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 attributes related to the training data used to generate the NN formation configuration. For example, input characteristics may include edge computing participation patterns (e.g., addition, omission, aggregation), cloud-based computing participation patterns (e.g., addition, omission, aggregation), power information, SINR information, CQI, CSI, Doppler feedback, RSS, error metric, minimum end-to-end (E2E) latency, expected E2E latency, E2E QoS, E2E throughput, E2E packet loss rate, service cost, etc.

[0051] CRM 306 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 transmitted via E2E communication, such as determining the E2E ML configuration based on one or more endpoint participation patterns as further described. Alternatively or additionally, the E2E ML controller analyzes any combination of the ML capabilities of the devices participating in the E2E communication (e.g., supported ML architectures, supported number of layers, available processing power, memory limitations, available power budget, fixed-point and floating-point processing, maximum kernel size capability, compute capability). In some implementations, the E2E ML controller obtains metrics characterizing the current operating environment and analyzes the current operating environment to determine the E2E ML configuration. This includes determining an E2E ML configuration that includes a combination of architecture configuration and parameter configuration(s) defining the DNN, or determining an E2E ML configuration that simply includes parameter configurations for updating the DNN.

[0052] When determining the E2E ML configuration, the E2E ML controller sometimes determines the partitioning of the E2E ML configuration, which distributes the processing functions associated with the E2E ML configuration across multiple devices. For clarity, Figure 3 The E2E ML controller 318 is illustrated as separate from the core network neural network manager 312; however, in alternative or other embodiments, the core network neural network manager 312 includes functions performed by the E2E ML controller 318, or vice versa. Furthermore, although Figure 3 The illustration shows the core network server 302 that implements the E2E ML controller 318, but alternatively or additionally, the device can implement the E2E ML controller, such as the base station 120 and / or other network elements.

[0053] The core network server 302 also includes a core network interface 320 for communicating user plane data, control plane information, and other data / information with other functions or entities in the core network 150, base station 120, ECS 180, or UE 110. In one implementation, the core network server 302 uses the core network interface 320 to transmit E2E ML configuration or portions of partitionable E2E ML configuration to the base station 120. Alternatively or additionally, the core network server 302 uses the core network interface 320 to receive feedback from the base station 120 and / or UE 110 via the base station 120.

[0054] ECS 180 includes one or more processors 322 and computer-readable storage medium 324 (CRM 324). Processor 322 may be a single-core or multi-core processor composed of various materials such as silicon, polysilicon, high-k dielectrics, copper, etc. CRM 324 may include any suitable memory or storage device for storing device data 326 of ECS 180, such as random access memory (RAM), static RAM (SRAM), dynamic RAM (DRAM), non-volatile RAM (NVRAM), read-only memory (ROM), hard disk drive, or flash memory. CRM 324 includes applications 328 and application data 330 used by the operating system of UE 110 and / or ECS 180, which may be executed by one or more processors 322 to enable communication with UE 110, base station 120, and core network server 302.

[0055] ECS 180 also includes an Xe interface 332 for communicating with base station 120 and a core network interface 334 for communicating with user plane data and / or control plane information with core network server 302.

[0056] Configurable machine learning module

[0057] Figure 4 The illustration depicts an example operating environment 400 including a UE 110 and a base station 120, encompassing various aspects capable of enabling E2E DNN adaptation for edge computing. In this implementation, the UE 110 and base station 120 exchange communications with each other on a wireless communication system by using multiple DNNs to process communications.

[0058] The base station neural network manager 268 of base station 120 includes a downlink processing module 402 for processing downlink communications, such as those for generating downlink communications to be transmitted to UE 110. For illustration, the base station neural network manager 268 forms one or more deep neural networks 404 (DNN 404) in the downlink processing module 402 using an E2E ML configuration and / or a portion of the E2E ML configuration, as further described. In various aspects, the DNN 404 performs some or all of the transmitter processing chain functions for generating downlink communications, such as receiving input data, proceeding to the encoding stage, followed by the modulation stage, and then the radio frequency (RF) analog transmission (Tx) stage. For illustration, the DNN 404 is capable of performing convolutional coding, serial-to-parallel conversion, cyclic prefix insertion, channel coding, time / frequency interleaving, etc. In some aspects, the DNN 404 processes edge computing communications, cloud-based computing communications, or any combination thereof.

[0059] Similarly, the UE neural network manager 218 of UE 110 includes a downlink processing module 406, wherein the downlink processing module 406 includes one or more deep neural networks 408 (DNN 408) for processing (received) downlink communication. In various embodiments, the UE neural network manager 218 uses an E2E ML configuration and / or a portion of an E2E ML configuration to form the DNN 408, as further described. In various aspects, the DNN 408 performs some or all of the receiver processing functions for (received) downlink communication, such as complementary processing to the processing performed by the DNN 404 (e.g., RF analog reception (Rx) phase, demodulation phase, decoding phase). For illustration, the DNN 408 is capable of performing any combination of tasks such as extracting data embedded in the Rx signal, recovering binary data, correcting data errors based on forward error correction applied in the transmitter frame, extracting payload data from frames and / or time slots, etc.

[0060] Base station 120 and / or UE 110 also use a DNN to process uplink communication. In environment 400, UE neural network manager 218 includes uplink processing module 410, wherein uplink processing module 410 includes one or more deep neural networks 412 (DNN 412) for generating and / or processing uplink communication (e.g., encoding, modulation). In other words, uplink processing module 410 processes pre-transmitted communication as part of processing uplink communication. UE neural network manager 218 forms DNN 412, for example, using E2E ML configuration and / or a portion of E2E ML configuration, to perform some or all of the transmitter processing functions for generating uplink communication transmitted from EE 110 to base station 120. In various aspects, DNN 412 performs some or all of the transmitter processing chain functions for generating uplink communication, such as receiving input data, proceeding to the encoding stage, followed by the modulation stage, and then the radio frequency (RF) analog transmission (Tx) stage.

[0061] Similarly, the uplink processing module 414 of base station 120 includes one or more deep neural networks 416 (DNN 416) for processing (received) uplink communication, wherein the base station neural network manager 268 uses an E2E ML configuration and / or a portion of an E2E ML configuration to form the DNN 416 to perform some or all of the receiver processing functions for (received) uplink communication, such as uplink communication received from UE 110. Sometimes, DNN 412 and DNN 416 perform complementary functions to each other. For example, DNN 416 may perform some or all of the receiver processing chain functions for (received) uplink communication, such as processing complementary to the processing performed by DNN 412 (e.g., RF analog reception (Rx) stage, demodulation stage, decoding stage).

[0062] Typically, a deep neural network (DNN) corresponds to a group of connected nodes organized into four or more layers. The nodes between layers can be configured in various ways, such as a partially connected configuration where a first subset of nodes in the first layer connects to a second subset of nodes in the second layer, or a fully connected configuration where every node in the first layer connects to every node in the second layer. Nodes can use various algorithms and / or analyses to generate output information based on adaptive learning, such as unilinear regression, multiple linear regression, logistic regression, stepwise regression, binary classification, multi-class classification, multivariate adaptive regression splines, local estimation scatter plot smoothing, etc. Sometimes, one or more algorithms include weights and / or coefficients that change based on adaptive learning. Therefore, the weights and / or coefficients reflect the information learned by the neural network.

[0063] Neural networks can also employ various architectures that determine which nodes within the network are connected, how data is advanced and / or retained within the network, what weights and coefficients are used to process the input data, and how the data is processed, among other things. These various factors collectively describe the NN formation configuration. For illustration, recurrent neural networks, such as Long Short-Term Memory (LSTM) neural networks, form loops between node connections to retain information from the first part of the input data sequence. The recurrent neural network then uses the retained information for subsequent parts of the input data sequence. As another example, feedforward neural networks pass information to the forward connections without forming loops to retain information. Although described in the context of node connections, it will be understood that the NN formation configuration can include various parameter configurations that influence how the neural network processes the input data.

[0064] The NN formation configuration of a neural network can be characterized by various architectures and / or parameter configurations. For illustration, consider an example of a DNN implementing a convolutional neural network. Typically, a convolutional neural network corresponds to a type of DNN where layers use convolution operations to process data to filter the input data. Accordingly, the convolutional NN formation configuration can be characterized, for example but not limited to, using pooling parameters (e.g., specifying pooling layers to reduce the dimensionality of the input data), kernel parameters (e.g., filter size and / or kernel type used in processing the input data), weights (e.g., biases used to classify the input data), and / or layer parameters (e.g., layer connections and / or layer types). While described in the context of pooling parameters, kernel parameters, weight parameters, and layer parameters, other parameter configurations can be used to form a DNN. Therefore, the NN formation configuration can include any other type of parameters that can be applied to the DNN to influence how the DNN processes the input data to generate output data. An E2E ML configuration uses one or more NN formation configurations to form an E2E DNN that handles communication from one endpoint to another. For example, a partitionable E2E ML configuration can be configured using a corresponding NameNode for each partition.

[0065] Figure 5 The illustration shows example 500, which generates multiple neural network (NN) configurations based on an adaptive description of an E2E DNN for edge computing. Sometimes, the aspects of example 500 are derived from... Figure 2 and Figure 3 It can be implemented by any combination of training module 270, base station neural network manager 268, core network neural network manager 312 and / or training module 314.

[0066] Figure 5The upper part includes DNN 502, which represents any suitable DNN for implementing E2E DNN adaptation for edge computing. In implementations, the neural network manager determines to generate different NN formation configurations, such as NN formation configurations for different operational configurations based on endpoint participation patterns (e.g., a first configuration adding ECS ​​to E2E communication, a second configuration omitting ECS ​​from E2E communication, a third configuration aggregating edge computing communication with cloud-based computing communication, and a fourth configuration excluding remote services from E2E communication). Alternatively or additionally, the neural network generates NN formation configurations based on different transmission environments and / or transmission channel conditions. Training data 504 represents example inputs to DNN 502, such as data corresponding to downlink and / or uplink communication with a specific operational configuration and / or a specific transmission environment. For illustration, training data 504 can include digital samples of downlink wireless signals, recovered symbols, recovered frame data, binary data, etc. In some implementations, the training module mathematically generates training data or accesses a file storing training data. Other times, the training module obtains real-world communication data. Therefore, the training module can train the DNN using mathematically generated data, static data, and / or real-world data 502. Some implementations generate input characteristics 506 that describe various qualities of the training data, such as operating configuration, transmit channel metrics, UE capabilities, UE speed, including edge computing, etc.

[0067] DNN 502 analyzes the training data and generates output 508, represented here as binary data. Some implementations iteratively train DNN 502 using the same training dataset and / or additional training data with the same input characteristics 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 some point in the training, such as when the training module determines that the accuracy meets or exceeds a desired threshold, or that the training process meets or exceeds the number of iterations, the training module determines to extract the architecture and / or parameter configuration 510 of the neural network (e.g., one or more pooling parameters, one or more kernel parameters, one or more layer parameters, weights). The training module then extracts the architecture and / or parameter configuration from the machine learning module to use as the NN formation configuration and / or one or more NN formation configuration elements. The architecture and / or parameter configuration can include any combination of fixed architecture and / or parameter configurations and / or variable architecture and / or parameter configurations.

[0068] Figure 5 The lower part includes a neural network table 512 representing a set of configuration elements for the NN, such as... Figure 2 and Figure 3The neural network tables 216, 272, and / or 316 are used. Neural network table 512 stores various combinations of architecture configurations, parameter configurations, and input characteristics; however, alternative implementations omit input characteristics from the table. Various implementations update and / or maintain NN formation configuration elements and / or input characteristics as the DNN learns additional information. For example, at index 514, the neural network manager and / or training module updates neural network table 512 to include architecture and / or parameter configurations 510 generated by the DNN 502 during the analysis of training data 504. At later points in time, such as when determining the E2E ML configuration for handling E2E communications involving added or omitted ECS endpoints, the neural network manager selects one or more NN formation configurations from neural network table 512 by matching input characteristics with the current operating environment and / or configuration, such as by matching input characteristics with current channel conditions, edge computing participation modes, UE capabilities, UE characteristics (e.g., speed, location, etc.).

[0069] E2E Architecture Adaptation for Edge Computing

[0070] End-to-end communication (E2E communication) involves two or more endpoints passing communication between each other, such as data sessions between UEs, remote services, and / or ECSs. E2E communication can correspond to unidirectional transmission, where a first endpoint sends communication and a second endpoint receives and / or resumes communication (e.g., downlink only, uplink only), or bidirectional transmission, where two endpoints send and receive communication with each other (e.g., reciprocity of downlink and uplink).

[0071] Various factors influence the configuration and operational efficiency of E2E communication and how network devices process and / or route communication via E2E. As an example, an ECS endpoint processes UE application data, and based on its closer proximity to the UE, the ECS endpoint delivers application data via E2E communication with less latency compared to cloud-based computing endpoints (e.g., Remote Service 170). As another example, the current operating environment (e.g., current channel conditions, UE location, UE movement, UE capabilities) affects the accuracy of data recovery by the receiving endpoint (e.g., bit error rate, packet loss). To illustrate, E2E communication using 5G millimeter wave (mmW) technology becomes more susceptible to signal distortion compared to lower frequency sub-6GHz signals. As yet another example, devices participating in E2E communication often have different capabilities and resources (e.g., memory storage, processor power). Adaptive E2E DNNs provide a flexible solution to the dynamic and changing factors affecting the performance of data delivery and / or recovery via E2E communication (e.g., higher processing resolution, faster processing, lower bit error rate, improved signal quality, improved latency).

[0072] In terms of E2E DNN adaptation for edge computing, the UE, base station, and / or core network server dynamically adapt and / or switch the E2E DNN's E2E ML configuration (e.g., architecture, parameters) based on the endpoint(s) participating in the E2E communication. For example, when E2E communication switches between an ECS and a remote service, the various aspects adapt the E2E ML configuration to the endpoint of the data session. To illustrate, a first E2E ML configuration directed to E2E communication involving an ECS can prioritize processing quality over computation latency. In other words, because ECS communication has lower delivery latency relative to other endpoints, the E2E ML configuration forms an E2E DNN that includes more processing layers and / or processes data at higher resolution to improve the quality of E2E communication. Alternatively, a second E2E ML configuration directed to E2E communication involving a remote server prioritizes computation latency over processing quality based on the higher delivery latency of the remote service relative to other endpoints. Therefore, the second E2E ML configuration, guided to handle communication with remote services, forms an E2E DNN with fewer processing layers, more data downsampling, a reduced number of fully connected layers, and / or lower processing resolution compared to the first E2E ML configuration. Various aspects are adaptively and / or reconfigured in the E2EML configuration to form an E2E DNN that aggregates and / or splits communication via E2E communication between multiple endpoints.

[0073] Figure 6 The illustration shows an example environment 600 that enables E2E DNN adaptation for edge computing based on various aspects. Environment 600 includes... Figure 1 UE 110, base station 120, remote service 170, ECS 180, and Figure 3 The core network server 302. In various aspects, the UE 110, base station 120, and / or core network server 302 implement sample partitions of the E2E DNN for processing information and / or data transmission over E2E communication.

[0074] For reference Figure 3The E2E ML controller 318 determines one or more E2E ML configurations that form one or more E2E DNNs for processing communications transmitted via one or more E2E communications. This includes determining adjustments to: (a) existing E2E ML configurations, such as minor adjustments using parameter updates (e.g., coefficients, weights) to tune one or more E2E DNNs based on feedback; and / or (b) ML architecture changes (e.g., number of layers, layer computation patterns (e.g., downsampling configuration, adding or removing fully convolutional layers), node connections) to reconfigure one or more E2E DNNs. In environment 600, the E2EML controller 318 determines an E2E ML configuration that forms a bidirectional E2E DNN, but in alternative or additional implementations, the E2EML controller 318 determines an E2E ML configuration that forms one or more unidirectional E2E DNNs.

[0075] like Figure 6 As shown, the E2E ML controller 318 determines a first E2E ML configuration for forming a first E2E DNN for processing communications transmitted via the first E2E communication 602 between UE 110 and remote service 170, and a second E2E ML configuration for forming a second E2E DNN for processing communications transmitted via the second E2E communication 604 between UE 110 and ECS 180. In various aspects, in response to changes in the participation of endpoints in E2E communications (e.g., changes in endpoints in a data session), such as when UE 110 moves into and out of the coverage area of ​​a base station (e.g., base station 120) connected to ECS (e.g., ECS 180), the E2E ML controller 318 dynamically determines different E2E ML configurations and adaptations to those configurations.

[0076] Because E2E communication 602 uses a different endpoint than E2E communication 604, E2E ML controller 318 determines different architectures and / or parameters for each E2E ML configuration based on different priorities. For example, in response to determining that existing E2E communication is modified by including remote service 170 as an endpoint, E2E ML controller 318 determines a first ML architecture based on prioritizing delivery latency over (e.g., minimizing delivery latency) processing quality (e.g., processing resolution, data frame rate). For example, E2E ML controller 318 selects an architecture with fewer processing layers and / or more data downsampling to reduce computation time, as further described. In response to determining that remote service 170 as an endpoint is removed and ECS 180 is added as an endpoint, E2E ML controller 318 identifies a second ML architecture based on prioritizing processing quality (e.g., higher processing resolution, higher frame rate) over processing latency, such as by selecting an ML architecture with more processing layers (e.g., more convolutional layers, fewer downsampling, etc.), which increases processing accuracy at the expense of increased computation time. Therefore, the E2E ML controller 318 dynamically adapts the E2E ML configuration based on changes in the participation mode of ECS and / or cloud-based services in E2E communication.

[0077] The E2E ML controller 318 can also determine the E2E ML configuration based on other factors, such as the machine learning (ML) capabilities of the devices or network entities involved in the E2E communication (e.g., supported ML architectures, supported layers, available processing power, memory limitations, available power budget, fixed-point processing versus floating-point processing, maximum kernel size capability, and computing power). As another example, the E2E ML controller analyzes the current operating environment, for instance, by analyzing link quality indications received from UE 110 and / or base station 120. In determining the E2E ML configuration, some implementations of the E2E ML controller partition the E2E ML configuration based on the devices involved in the E2E communication and transmit the appropriate partition of the E2E ML configuration to each appropriate device. As an example, the core network server 302 sends messages, such as Non-Access Stratum (NAS) messages, to UE 110 to indicate architecture and / or parameter changes to the DNN implemented by the UE that processes data routed through E2E communication.

[0078] E2E ML controller 318 determines a first E2E ML configuration for forming a first E2E DNN to process communications transmitted via E2E communication 602. E2E ML controller 318 uses any combination of information to identify the first E2E ML configuration, such as endpoint participation modes (e.g., adding remote service 170, omitting ECS ​​180), endpoint characteristics (e.g., latency, throughput), priority, link quality indications, performance requirements (e.g., resource type, priority, packet delay budget, packet error rate, maximum data burst size, average window, security level), available radio network resources, ML capabilities of participating devices (e.g., base station 120, core network server 302, UE 110), current operating environment (e.g., channel conditions, UE location), etc.

[0079] In various aspects, the E2E ML controller 318 identifies characteristics of the remote service 170 by analyzing historical records and / or metrics. For illustration, the E2E controller 318 accesses historical records indicating the statistical round-trip latency characteristics of data transfers between the core network server 302 and the remote service 170. If the latency characteristics exceed a performance threshold or consume a significant portion of the time budget allocated to data transfer, the E2E ML controller 318 selects an E2E ML configuration that is directed to handle data transfers via E2E communication relative to other architectures within a predetermined (shorter) time frame to offset or compensate for the remote service's latency characteristics and / or maintain the time budget. For illustration, an end-to-end round-trip latency time budget of 100 milliseconds (msec) is considered. In response to the determination that the round-trip latency characteristics of communication between the core network server 302 and the remote service 170 statistically exceed 50 milliseconds (and thus occupy most of the round-trip latency time budget), the E2E ML controller 318 identifies a first E2E ML configuration with a first architecture that processes data transfers via E2E communication at a lower resolution within 10 milliseconds compared to a second architecture that processes data transfers via E2E communication within 20 milliseconds at a higher resolution. For example, the E2E ML controller 318 selects a first E2E ML configuration relative to the second E2E ML configuration that has fewer layers, alters layer computation modes to increase or decrease downsampling performed by layers, fewer fully connected layers, and / or fewer processing nodes.

[0080] The E2E ML controller 318 sometimes determines partitions of the E2E ML configuration (and the resulting E2E DNN formed using the E2E ML configuration) to distribute processing among the various devices participating in the E2E communication. In other words, the E2E ML configuration forms a distributed E2E DNN, in which multiple devices implement corresponding portions of the distributed E2E DNN. For example, in response to determining a first E2E ML configuration associated with E2E communication 602, the E2E ML controller 318 partitions the first E2E ML configuration into multiple portions and directs the devices to form corresponding DNNs based on these portions. For illustration, the core network server 302 uses a first portion of the first E2E ML configuration to form a first (central) DNN 606 to handle communication transmissions via the core network server through the Internet 160 to remote service 170 (and / or other cloud-based services) for E2E communication 602 (and / or other cloud-based services). In some aspects, the E2E ML controller 318 may alternatively or additionally partition the first E2E ML configuration to form a DNN (e.g., a portion of the E2E DNN) at a data center (not shown). In various aspects, the E2E ML controller 318 partitions the E2E DNN to offload compute-intensive operations at base station 120 (e.g., via DNN 606) to the core network server. Alternatively or additionally, DNN 606 routes communications to and from remote services via the Internet, such as by generating network packets to send to remote service 170 via the Internet and / or by receiving network packets from remote service 170 in a manner optimized for E2E communications.

[0081] Similarly, base station 120 uses a second portion of a first E2E ML configuration for processing communications transmitted via E2E communication 602 to form a second DNN 608, wherein DNN 606 and DNN 608 communicate with each other at interface 610 as part of E2E communication 602. In other words, DNN 606 provides input to DNN 608, and / or DNN 608 provides input to DNN 606. UE 110 uses a third portion and / or partition of the first E2E ML configuration to form a third DNN 612 to process communications transmitted via E2E communication 602. Collectively, DNNs 606, 608, and 612 correspond to distributed (and partitionable) E2E DNNs formed by the E2E ML configuration, wherein DNN 606 typically corresponds to a central DNN that is directed to process (via Internet 160) communications with remote service 170.

[0082] In various aspects, the core network server 302, base station 120, and / or UE 110 determine to modify E2E communication 602 to add ECS 180 and omit remote service 170. In other words, the core network server 302, base station 120, and / or UE 110 determine to adaptively and / or change the E2E DNN associated with E2E communication 602 to form the E2E DNN associated with E2E communication 604. However, as referenced... Figure 7 As described, the core network server 302, base station 120, and / or UE 110 sometimes modify the E2E ML configuration to form an E2E DNN that handles input / output with multiple endpoints. The core network server 302, base station 120, and / or UE 110 are capable of determining modifications to E2E communication based on various factors. As an example, UE 110 indicates its estimated UE location to the core network server via base station 120, and the core network server 302 determines, based on delivery latency, to switch UE 110 from the data server to an ECS within a predetermined distance of the estimated UE location. As another example, UE 110 sends a request to base station 120 to add an ECS to increase data throughput and / or improve latency. As yet another example, base station 120 identifies the specific ECS closest to the UE based on the estimated UE location (from among multiple ECSs connected to the base station).

[0083] In response to determining that mobile edge computing will be included in E2E communication, the E2E ML controller 318 determines a second E2E ML configuration, wherein the second E2E ML configuration can correspond to an E2E ML configuration that is separate from and different from the first E2E ML configuration, or can correspond to adjustments, tuning, refinements, and / or architectural updates to the first E2E ML configuration. The E2E ML controller 318 uses any combination of information (e.g., endpoint participation mode, endpoint characteristics, link quality indications, performance requirements) to determine the second E2E ML configuration. As an example, the E2E ML controller 318 determines that the round-trip latency characteristics between ECS 180 and base station 120 (and / or core network server 302) statistically occur at 20 milliseconds or less. Again assuming a round-trip time budget of 100 milliseconds, the E2E ML controller 318 selects a second E2E ML configuration with an ML architecture that includes more layers and / or nodes compared to the first E2E ML configuration, because the round-trip time characteristics allow for more processing by the corresponding E2E DNN (e.g., higher resolution, higher frame rate).

[0084] In various aspects, the E2E ML controller 318 partitions the second E2E ML configuration into multiple parts and guides participating devices to form a DNN for processing communications transmitted via E2E communication 604. For example... Figure 6As shown, base station 120 forms a fourth (local) DNN 614 based on the first partition of the second E2E ML configuration, wherein DNN 614 communicates with ECS 180 at interface 191, as further described. However, in an alternative or additional implementation, core network server 302 forms the fourth (local) DNN 614 and communicates with ECS 180 at interface 192 ( Figure 6 (Not shown in the image) communicates with ECS 180. In various aspects, such as by generating communication packets to ECS 180 and / or by receiving communication from ECS 180 in a format optimized for E2E communication, DNN 614 routes communication between base station 120 and ECS 180.

[0085] The E2E ML controller 318 also updates the second DNN 608 based on the second partition of the second E2E ML configuration, and updates the third DNN 612 based on the third partition of the second E2E ML configuration. In E2E communication 604, DNN 608 receives input from local DNN 614 at interface 616 instead of receiving input from DNN 606 at interface 610, as illustrated for E2E communication 602. The partitioning representation of the second E2E ML configuration shown by the DNN in E2E communication 604 is an example partition, and the E2E controller 318 is capable of partitioning and distributing the E2E ML configuration (and corresponding E2E DNNs) in other ways.

[0086] For clarity, UE 110 and base station 120 update and use the same DNN (e.g., DNN 612, DNN 608) for both E2E communication 602 and E2E communication 604. However, in alternative or other implementations, UE 110 and base station 120 maintain separate DNNs for different E2E communications.

[0087] Figure 7 The illustration shows an example environment 700 that enables E2E DNN adaptation for edge computing based on various aspects. Environment 700 includes... Figure 1 UE 110, base station 120, remote service 170 and ECS 180, and Figure 3 The core network server 302.

[0088] In environment 700, UE 110 acts as the first endpoint of E2E communication 702, remote service 170 acts as the second endpoint, and ECS 180 acts as the third endpoint. In each aspect, E2E ML controller 318 determines an E2E ML configuration for forming an E2E DNN that aggregates and / or splits communications transmitted between UE 110 and the two other endpoints. While E2E ML controller 318 determines an E2E ML configuration for a bidirectional E2E DNN, E2E ML controller 318 is also capable of alternatively or additionally determining one or more E2E ML configurations for a unidirectional DNN to aggregate and / or split downlink or uplink communications, as further described.

[0089] UE 110, base station 120, and / or core network server 302 form a distributed E2E DNN to process information and / or data transmitted via E2E communication 702, wherein UE 110 uses both remote service 170 and ECS 180 to process application data. In various aspects, E2E ML controller 318 determines the E2E ML configuration for the distributed E2E DNN, which splits communication (via Internet 160) from UE 110 to remote service 170 and ECS 180, and then aggregates communication (via Internet 160) from remote service 170 and ECS 180 to UE 110. Similar to reference [reference missing]. Figure 6 The E2E ML controller 318 uses any combination of information such as priority ordering, endpoint participation mode, endpoint characteristics, link quality indication, performance requirements, available wireless network resources, ML capabilities of participating devices, and current operating environment to identify the E2E ML configuration.

[0090] As an example, the E2E ML controller 318 analyzes device capabilities and guides the UE to form a DNN with fewer layers and a smaller kernel size relative to the 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) with an architecture that processes information without exceeding the UE's memory constraints. In some instances, the E2E ML controller calculates whether the corresponding computational load performed at each device collectively meets the performance requirements corresponding to the latency budget and determines the E2E ML configuration designed to meet the performance requirements.

[0091] Environment 700 illustrates an example partition where the E2E ML controller 318 divides the E2E ML configuration (and the E2E DNN formed using the E2E ML configuration) into five parts. The core network server 302 uses the first part of the E2E ML configuration to form a first DNN 704. The base station 120 uses the second part of the E2E ML configuration to form a second DNN 706, the third part to form a third DNN 708, and the fourth part to form a fourth DNN 710. The UE 110 uses the fifth part of the E2E ML configuration to form a fifth DNN 712. The partitions and functions illustrated by DNNs 704, 706, 708, 710, and 712 represent the functions of the distributed E2E DNN and the example partitions formed. In an alternative implementation, the E2E ML controller 318 partitions the E2E ML configuration in other ways. As an example, the E2E ML controller 318 partitions the E2E ML configuration to combine the processes described for DNNs 706, 708, and 710 into a single DNN implemented at base station 120. As a second example, the E2E ML controller 318 partitions the E2E ML configuration to include an additional DNN at core network server 302, which is located at interface 192 (…). Figure 7 Inputs are received from ECS 180 (not shown in the diagram), such as those for mobility management associated with edge computing. For illustration, E2E ML controller 318 partitions the E2E DNN to offload compute-intensive operations at base station 120 to the core network server (e.g., via DNN 704). Alternatively or additionally, DNN 704 routes communications to and from remote services via the Internet, such as by generating network packets to send to remote service 170 via the Internet and / or by receiving network packets from remote service 170 in a manner optimized for E2E communications. Similarly, in some aspects, DNN 706 routes communications between base station 120 and ECS 180, such as by generating communication packets to ECS 180 and / or by receiving communications from ECS 180 using a format optimized for E2E communications.

[0092] For downlink communication, DNN 704 receives a first input (e.g., application data) from remote service 170, processes the first input, and generates a first output. Similarly, DNN 706 receives a second input from ECS 180 at interface 191, processes the second input, and generates a second output. DNN 708 receives and processes the first and second outputs from the respective endpoints and generates an aggregated output received and processed by DNN 710. DNN 710 then transmits the corresponding result to DNN 712.

[0093] For uplink communication, DNN 712 at UE 110 generates an output that DNN 710 uses as input. Based on this input, DNN 710 at base station 120 generates a single output received by DNN 708, which is also implemented at base station 120. DNN 708 generates a split output: a first output directed to DNN 704 at core network server 302 (and subsequently to remote service 170 via Internet 160), and a second output directed to DNN 706 (and subsequently to ECS 180 at interface 191).

[0094] Determining the E2E ML configuration and adaptively configuring the E2E ML based on one or more corresponding endpoints allows network entities to dynamically modify the E2E DNN processing communication as the UE moves and the corresponding E2E communication endpoints change. In some aspects, network entities determine a partitionable E2E ML configuration to distribute E2E DNN processing and / or direct devices with fewer resources to form a DNN with less processing (e.g., less data, less memory, fewer CPU cycles, fewer nodes, fewer layers) compared to devices with more processing resources and / or memory. Dynamic adaptation and / or partitioning allows network entities to modify the (distributed) E2E DNN based on the endpoints participating in E2E communication and to improve E2E communication performance by referencing one or more metrics, such as higher resolution, faster processing, lower bit error rate, improved signal quality, improved latency, etc.

[0095] E2E DNN Adaptive Signaling and Control Transactions for Edge Computing

[0096] Figure 8 and 9 The diagram illustrates an example signaling and control transaction between core network servers, base stations, and user equipment, adapted to one or more aspects of an E2E DNN for edge computing. The operations of the signaling and control transactions can be performed by... Figure 1 Base station 120 and UE 110 or Figure 3 Core network server 302 usage as per reference Figure 1-7 To perform any of the described aspects.

[0097] The first example of E2E DNN adaptive signaling and control transactions for edge computing is provided by Figure 8 The signaling and control transaction diagram 800 is illustrated. As shown in 805, the core network server 302, base station 120, and UE 110 use an E2E DNN to process communications transmitted via E2E communication. In various aspects, the E2E DNN corresponds to a distributed E2E DNN, such as reference... Figure 6 and 7Those described.

[0098] In 810, the core network server 302 determines to modify the E2EDNN based on changes in the participation mode of the ECS in E2E communication. As an example, the core network server 302 receives requests or notifications from UE110 and / or base station 120 to add an ECS to E2E communication for application processing, such as... Figure 9 As further described in [the text]. Therefore, sub-Figure 815 generally represents signaling and control transactions that can be used to determine when and how to modify the E2E DNN based on changes in endpoint participation modes, and can include various signaling and control transactions between core network server 302, base station 120, and / or UE 110. The determination of modifying the E2E DNN can be based on changes such as adding and utilizing ECS, as described in E2E communication 604; based on omitting ECS, as described in E2E communication 602; or on changes such as aggregating communication with ECS and communication with remote services, as described in E2E communication 702.

[0099] At 820, the core network server 302 identifies or determines the E2E ML configuration based on changes in the ECS participation mode via the E2E ML controller 318. For illustration, the core network server 302 analyzes the performance characteristics (e.g., latency characteristics), prioritization, and / or performance requirements (e.g., latency budget) of the E2E communication based on the endpoints included in the change. Based on the ECS participation mode, the E2E ML controller 318 determines the E2E ML architecture, such as the number of processing layers included, the computation mode and / or configuration of each layer (e.g., downsampling configuration, number of connected nodes, number of convolutional layers).

[0100] Alternatively or additionally, the E2E ML controller 318 analyzes the capabilities of the devices participating in E2E communication, such as UE capabilities. Sometimes, the E2E ML controller analyzes the radio network resource partitions associated with the E2E communication. As another example, the E2E ML controller 318 analyzes metrics characterizing the current operating environment and determines the E2E ML configuration based on the current operating environment. In one implementation, the E2E ML configuration corresponds to unidirectional E2E communication for uplink or downlink data services, while in other implementations, the E2E communication corresponds to bidirectional E2E communication for both uplink and downlink data services.

[0101] When identifying E2E ML configurations, core network server 302 sometimes analyzes one or more neural network tables to obtain one or more neural network-formed configurations corresponding to the E2E ML configuration. For example, the core network server partitions E2E ML configurations across multiple devices, as shown in the reference. Figure 6 and 7As described, and for each partition, a corresponding entry is determined in the neural network table, where each entry indicates the architecture and / or parameter configuration. Therefore, the core network server 302 is able to determine the distributed E2E ML configuration that forms the distributed E2E DNN, and determine the partitions of the E2E ML configuration / E2E DNN for devices participating in E2E communication.

[0102] At 825, the core network server 302 bootstraps the device to form the DNN based on the E2E ML configuration identified at 820. For example, as referenced... Figure 6 and Figure 7 As described, the core network server 302 instructs the base station 120 to form and / or update a first part of the E2E DNN using a first part of the E2E ML configuration, and instructs the UE 110 to form and / or update a second part of the E2E DNN using a second part of the E2E ML configuration. In some embodiments, the core network server 302 instructs the device to form the corresponding DNN by indicating indexes to neural network tables, such as neural network table 216 and / or neural network table 272. Sometimes, the base station 120 forwards this instruction from the core network server 302 to the UE 110, as shown at 830.

[0103] At 835, core network server 302 optionally forms a DNN (e.g., DNN 606, DNN 704) based on the E2E ML configuration identified at 820. In an implementation, the DNN formed by the core network server performs at least some processing of transmitting communications via E2E communication, such as application data to / from remote service 170. Similarly, at 840, base station 120 forms and / or updates a DNN (e.g., DNN 608, DNN 614, DNN 706, DNN 708, DNN 710) based on the E2E ML configuration identified at 820. For example, base station 120 accesses a neural network table to obtain, as referenced... Figure 5 One or more parameters and / or architectures are described. In an implementation, the DNN formed by base station 120 performs at least some of the following processes: transmitting application data via E2E communication, such as application data transmitted to / from ECS 180, aggregation and / or splitting of application data processed by ECS 180 and remote service 170, or application data transmitted to / from remote service 170. At 845, UE 110 forms and / or updates the DNN (e.g., DNN 612, DNN 712) based on the E2E ML configuration determined at 820. For example, UE 110 accesses a neural network table to obtain, as referenced... Figure 5One or more parameters and / or architectures are described. In the implementation, the DNN formed by UE 110 performs at least some of the processing to transmit information and / or data via E2E communication over the wireless network.

[0104] Subsequently, at 850, the core network server 302, base station 120, and / or UE 110 use the E2E DNN formed from the E2EML configuration identified at 820 to process communications transmitted via E2E communication. For example, refer to Figure 6 The DNN processes the uplink and / or downlink transmission of application data processed by remote server endpoints or ECS. As another example, see [reference needed]. Figure 7 DNNs handle the uplink and / or downlink transmission of application data by aggregating, splitting, and / or routing data associated with both remote service endpoints and ECS.

[0105] In various aspects, the core network server 302, base station 120, and / or UE 110 iteratively perform... Figure 8 The signaling and control transactions, represented by dashed line 855, are as described in Figure 800. These iterations allow the core network server 302, base station 120, and / or UE 110 to dynamically adjust and / or switch the DNN used for adaptive E2E communication as the participation mode of the ECS changes, such as based on changes in UE mobility. Adjustments can include architectural and / or parameter changes to the E2E DNN, as further described.

[0106] A second example of E2E DNN adaptive signaling and control transactions for edge computing is provided by Figure 9 The signaling and control transactions are illustrated in Figure 900. Figure 900 provides a diagram for executing... Figure 8 Example signaling and control transactions in subgraph 815.

[0107] At 905, UE 110 indicates the estimated UE location information to base station 120. As an example, UE 110 transmits a link quality indication to base station 120, and base station 120 generates the estimated UE location, such as through power level and / or timing information (e.g., time of arrival). As another example, the base station receives the estimated UE location from UE 110, such as by transmitting Global Positioning System (GPS) and / or Global Navigation Satellite System (GNSS) location information using a low-frequency channel (e.g., 700MHz, 800MHz).

[0108] At 910, base station 120 indicates the estimated UE location to core network server 302. Alternatively or additionally, base station 120 indicates to core network server 302 the applications invoked and / or executed at the UE, as further described at 920. In some aspects, base station 120 identifies the ECS server based on the estimated UE location and requests from core network server 302 to include or add the ECS server to E2E communication.

[0109] In response to receiving the request from base station 120, core network server 302 determines at 915 to modify the E2E DNN based on the estimated UE location. For illustration, it is assumed that existing E2E communication for UE 110 includes connections to data centers and / or remote services, as further described. Core network server 302 determines from the estimated UE location that the average round-trip time between the data center and / or remote service and the UE at the estimated UE location exceeds the round-trip time budget. For example, core network server 302 accesses historical data records that archive past round-trip times between other UEs at the estimated UE location and the data center. Alternatively or additionally, core network server 302 determines based on the estimated UE location that the UE is within a predetermined distance from the ECS, and / or determines the average round-trip time between the ECS and other UEs at the estimated UE location. In each aspect, core network server 302 determines the average round-trip time improvement associated with the ECS and then determines to add ECS 180 to the E2E communication.

[0110] In response to determining to add ECS 180 to E2E communication, core network server 302 determines to modify the E2E DNN, such as by forming an E2E DNN that routes application data to the ECS. Alternatively or additionally, core network server 302 determines to modify the E2E DNN to omit or exclude communication with the data center and / or remote services from E2E communication. In some aspects, core network server 302 identifies the E2E ML configuration that forms the E2E DNN to aggregate and / or split communication between the UE, data center, and ECS, as further described.

[0111] A third example of E2E DNN adaptive signaling and control transactions for edge computing, provided by Figure 9 The signaling and control transactions are illustrated in Figure 902. Figure 902 provides a diagram for executing... Figure 8 Example signaling and control transactions in subgraph 815.

[0112] At 920, UE 110 requests edge computing by transmitting a request to base station 120. For illustration, UE 110 identifies a call to a data-intensive application and / or an application with low latency requirements and requests mobile edge computing to serve that application. In some respects, the UE explicitly requests the addition of edge computing, while in others, the UE implicitly requests the addition of edge computing. For example, the UE implicitly requests the inclusion of edge computing by transmitting an indication of the invoked application and / or the data services utilized by that application.

[0113] At 925, in response to receiving a request from the UE to add edge computing, base station 120 requests an E2E DNN change from core network server 302. In some aspects, base station 120 determines, based on information indicated at 920, that the ECS connected to the base station supports processing for the invoked application and / or includes data services utilized by the application. Base station 120 then requests an E2E DNN change to add ECS 180 to the E2E communication. In other aspects, the base station determines that computation for the invoked application and / or data services has been completed at ECS 180 and requests an E2E DNN change to omit ECS 180 from the E2E communication. In response to receiving the request and / or in response to determining an authorization request, at 920, core network server 302 determines to modify the E2E DNN at 915, as further described.

[0114] Example Method

[0115] Based on one or more aspects of the E2E DNN adaptation used for edge computing, referencing Figure 10 and Figure 11 Example methods 1000 and 1100 are described. 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 can be skipped or combined in any order to implement the method or an alternative method. Generally, any components, modules, methods, and operations described herein can 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 computer-readable storage memory local and / or remote on a computer processing system, and implementations may include software applications, programs, functions, etc. Alternatively or additionally, any functionality described herein can be performed at least in part by one or more hardware logic components, such as, but not limited to, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-chips (SoCs), complex programmable logic devices (CPLDs), etc.

[0116] Figure 10An example method 1000 is illustrated for performing various aspects of E2E DNN adaptation for edge computing. In some implementations, the operation of method 1000 is performed by a network entity such as core network server 302.

[0117] At 1005, a network entity guides the UE and base station participating in end-to-end (E2E) communication to achieve E2E communication by forming at least a portion of an E2E deep neural network (DNN) based on a first E2E ML configuration. For example, the network entity (e.g., core network server 302) guides the base station (e.g., base station 120) to form a first DNN using a first portion of the E2E machine learning (ML) configuration, and guides the UE (e.g., UE 110) to form a second DNN using a second portion of the E2E ML configuration, as in... Figure 8 As described in 805.

[0118] In 1010, the network entity determines to update the first E2E ML configuration based on changes in the ECS participation mode during E2E communication. For example, the network entity (e.g., core network server 302) receives a request from the UE (e.g., UE 110) to add an ECS (e.g., ECS180), as in... Figure 9 As described in 925 and 930. As another example, a network entity (e.g., core network server 302) receives a request from a base station (e.g., base station 120) for a change in endpoint participation mode, as described in... Figure 9 As described in 915. For illustration, the network entity determines to update the first E2E ML configuration based on adding an ECS to E2E communications, aggregating communications with the ECS and communications with remote services in E2E communications, or removing a remote service from E2E communications. In some implementations, the network entity detects changes in the ECS's participation mode, such as by receiving a request from the UE to add edge computing to E2E communications or by receiving an indication from the base station that an ECS has been omitted from E2E communications.

[0119] In 1015, the network entity identifies the second E2E ML configuration based on the change in the ECS participation mode in E2E communication. For example, the network entity (e.g., core network server 302) identifies the second E2E ML configuration, as in... Figure 8As described in 820. When identifying a second E2E ML configuration, the network entity sometimes determines partitions that distribute the second E2E ML configuration across multiple devices. As an example, the network entity identifies a downlink E2E ML configuration as at least a portion of a second E2E ML configuration that forms a downlink E2E DNN, which is directed to receive a first input from a remote service, a second input from an ECS, and to aggregate the first and second inputs to generate an output directed to the UE. Alternatively or additionally, the network entity identifies an uplink E2E ML configuration that forms an uplink E2E DNN as at least a portion of the second E2E ML configuration, which is directed to receive input from the UE, using the input to generate a first output directed to the ECS as a first endpoint of E2E communication, and using the input to generate a second output directed to a remote service as a second endpoint of E2E communication. When identifying the second E2E ML configuration, the network entity identifies any combination of one or more parameter changes and / or architecture changes, such as changes to (one or more) coefficients, changes to multiple processing layers, changes to the computation mode of at least one processing layer, etc.

[0120] At 1020, the network entity guides at least the UE or base station to use a second E2E ML configuration to update at least a portion of the E2E DNN for enabling E2E communication. For example, the network entity (e.g., core network server 302) guides a base station (e.g., base station 120) to form a first DNN based on a first partition of the second ML configuration, and guides a UE (e.g., UE 110) to form a second DNN based on a second partition of the second ML configuration, as in... Figure 8 As described in 825. In at least one example, the network entity sends a NAS message to the UE to instruct for an update of the DNN.

[0121] In some aspects, method 1000 is repeated iteratively, as shown in 1025. For example, suppose that when a UE (e.g., UE 110) enters a coverage area that includes an ECS (e.g., ECS 180), the first iteration adds the ECS to the E2E communication. When the UE moves out of the coverage area, the second iteration removes the ECS from the E2E communication. This allows network entities to dynamically adapt the DNN and how the DNN handles E2E communication to optimize (and re-optimize) processing when endpoints in the E2E communication change.

[0122] Figure 11An example method 1100 for performing various aspects of E2E DNN adaptation for edge computing is illustrated. In some implementations, the operation of method 1100 is performed by a wireless transmit / receive unit (WTRU) such as a UE (e.g., UE 110) or a base station (e.g., BS 120).

[0123] In 1105, the WTRU uses at least a first portion of the first E2E ML configuration for implementing the E2E DNN to form the DNN. For example, the WTRU (e.g., UE 110) uses as in Figure 8 The first E2E ML configuration described in 805 is used to form the DNN (e.g., DNN 612, DNN 712). As another example, the WRTU (e.g., base station 120) uses at least a first portion of the first E2E ML configuration to form the DNN (e.g., DNN 608, DNN 614, DNN 706, DNN 708, DNN 710).

[0124] At 1110, the WTRU receives an instruction to update the DNN using at least a second portion of the second E2E ML configuration, based on a change in the ECS participation mode during E2E communication. For example, the WTRU (e.g., UE 110) receives an instruction to update the DNN (e.g., DNN 612, DNN 712), as in... Figure 8 As described in 830. In some respects, the UE receives an indication in the NAS message. As another example, the WTRU (e.g., base station 120) receives an indication to update the DNN (e.g., DNN 608, DNN 614, DNN 706, DNN 708, DNN 710), as in Figure 8 As described in 825.

[0125] In 1115, the WTRU uses at least the second part of the second E2E ML configuration to update the DNN. For example, the WTRU (e.g., UE 110) updates the DNN to form a DNN (e.g., DNN 612, DNN 712), which processes communication based on changes in the ECS participation mode in E2E communication, as in Figure 8 As described in 845. As another example, the WTRU (e.g., base station 120) updates the DNN to form a DNN (e.g., DNN 608, DNN 614, DNN 706, DNN 708, DNN 710), which processes communication based on changes in the participation mode of the ECS in E2E communication, as in... Figure 8 As described in 840.

[0126] In 1120, the WTRU uses the updated DNN to implement at least a portion of the E2E communication. For example, the WTRU (e.g., UE110, base station 120) uses the updated DNN to process communication transmitted via E2E communication, as in... Figure 8 As described in 850.

[0127] In some aspects, method 1100 is repeated iteratively, as shown in 1125. For example, suppose the first iteration forms a DNN that processes edge computation while the UE (e.g., UE 110) moves within a predetermined distance of the ECS (e.g., ECS 180). The second iteration removes edge computation when the UE moves away from the ECS and / or moves beyond the predetermined distance. This allows network entities to dynamically adapt the DNN and how the DNN processes communication to optimize (and re-optimize) processing when endpoints change in E2E communication.

[0128] Although techniques and apparatus for E2E DNN adaptation for edge computing have been described in feature- and / or method-specific language, it will be understood that the subject matter of the appended claims is not necessarily limited to the specific features or methods described. Rather, the specific features and methods are disclosed as exemplary embodiments of E2E DNN adaptation for edge computing.

[0129] Several examples are described below:

[0130] Example 1: A method for adaptive end-to-end E2E machine learning (ML) configuration performed by a network entity, the E2E ML configuration forming an E2E deep neural network (DNN) for processing communication transmitted via E2E communication between at least two endpoints using a wireless network, the method comprising: instructing a user equipment (UE) participating in the E2E communication to implement the E2E communication by forming at least a first portion of the E2E DNN based on a first E2E ML configuration; instructing a base station participating in the E2E communication to implement the E2E communication by forming at least a second portion of the E2E DNN based on the first E2E ML configuration; determining to update the first E2E ML configuration based on a change in the participation mode of an edge computing server (ECS) in the E2E communication; identifying a second E2E ML configuration based on a change in the participation mode of the ECS in the E2E communication; and instructing at least the UE or the base station to update at least a third portion of the E2E DNN using the second E2E ML configuration for implementing the E2E communication.

[0131] Example 2: According to the method of Example 1, wherein determining to update the first E2E ML configuration includes: determining to include the ECS in the E2E communication; and determining to update the first E2E ML configuration based on determining to include the ECS in the E2E communication.

[0132] Example 3: According to the method of Example 2, wherein determining to update the first E2E ML configuration based on determining that the ECS is included further comprises: determining to update the first E2E ML configuration based on: aggregating communication with the ECS and communication with the remote service in the E2E communication; or excluding the remote service from the E2E communication.

[0133] Example 4: According to the method of Example 3, wherein determining to update the first E2E ML configuration includes: determining to update the first E2E ML configuration based on the aggregation, and wherein identifying the second E2E ML configuration includes: identifying the downlink E2E ML configuration forming the downlink E2E DNN as at least a portion of the second E2E ML configuration, the downlink E2E DNN being directed to: receive a first portion of application data from the ECS; receive a second portion of the application data from the remote service; and aggregate the first portion and the second portion to generate aggregated application data directed to the UE.

[0134] Example 5: The method according to Example 3 or Example 4, wherein determining to update the first E2E ML configuration includes: determining to update the first E2E ML configuration based on the aggregation, and wherein identifying the second E2E ML configuration includes: identifying the uplink E2E ML configuration forming the uplink E2E DNN as at least a portion of the second E2E ML configuration, the uplink E2E DNN being directed to: receive uplink application data from the UE; generate a first output directed to the ECS using the uplink application data; and generate a second output directed to the remote service using the uplink application data. Example 6: The method according to any of the preceding examples, wherein determining to update the first E2E ML configuration further includes: receiving a request from the UE to include the ECS in the E2E communication; or determining to include the ECS in the E2E communication based on an estimated UE location.

[0135] Example 7: According to the method of Example 6, receiving the request from the UE to include the ECS in the E2E communication further includes: receiving an implicit request to include the ECS in the E2E communication.

[0136] Example 8: The method according to any of the foregoing examples, wherein guiding at least the UE or the base station to update the at least third portion of the E2E DNN using the second E2E ML configuration further comprises: guiding the UE to update the first portion of the E2E DNN using the second E2E ML configuration; or guiding the base station to update the second portion of the E2E DNN using the second E2E ML configuration.

[0137] Example 9: The method according to any of the preceding examples, wherein identifying the second E2E ML configuration includes at least one of the following: identifying one or more parameter changes of the E2E DNN; or identifying one or more architectural changes of the E2EDNN.

[0138] Example 10: According to the method of Example 9, wherein identifying the change of the one or more parameters further includes: identifying the change of one or more coefficients.

[0139] Example 11: The method according to any of the foregoing examples, wherein instructing at least the UE or the base station to update at least a third portion of the E2E DNN further comprises: sending a non-access stratum message to the UE indicating the updating of the at least third portion of the E2E DNN.

[0140] Example 12: A method for adaptive end-to-end E2E machine learning (ML) configuration executed by a wireless transmit / receive unit (WTRU) for processing communications transmitted via E2E communications in a wireless network, the method comprising: forming a deep neural network (DNN) using at least a first portion of a first E2E ML configuration for implementing the E2E DNN; receiving an instruction to update the DNN using at least a second portion of a second E2E ML configuration based on a change in the participation mode of an edge computing server (ECS) in the E2E communications; updating the DNN using at least a second portion of the second E2E ML configuration; and implementing at least a portion of the E2E communications using the updated DNN.

[0141] Example 13: The method according to Example 12 further includes: identifying the edge computing server based on the estimated UE location of the WTRU; and requesting that the ECS be included in the E2E communication.

[0142] Example 14: According to the method of Example 12 or Example 13, wherein receiving the instruction to update the DNN further includes: receiving guidance to update one or more parameters of the DNN as the instruction; or receiving guidance to update the architecture of the DNN as the instruction.

[0143] Example 15: According to the method of Example 14, receiving guidance to update the architecture further includes at least one of the following: changing the number of processing layers used in the DNN; and changing the computation mode of at least one processing layer in the DNN.

[0144] Example 16: According to the method of Example 15, changing the number of processing layers includes adding one or more convolutional layers to the DNN.

[0145] Example 17: The method according to any one of Examples 14 to 16, wherein receiving guidance to update the one or more parameters further comprises: receiving guidance to update one or more coefficients of the at least portion of the DNN.

[0146] Example 18: The method according to any one of Examples 12 to 17, wherein the WTRU includes a User Equipment (UE), and wherein receiving an instruction to update the DNN further includes: receiving the instruction from a base station.

[0147] Example 19: A network entity includes: a processor; and a computer-readable storage medium including instructions for implementing an end-to-end machine learning controller for performing the method according to any one of Examples 1 to 11.

[0148] Example 20: A user equipment includes: a processor; and a computer-readable storage medium including instructions that are executed by the processor to direct the user equipment to perform the method according to any one of Examples 12 to 18.

[0149] Example 21: A base station includes: a processor; and a computer-readable storage medium including instructions that are executed by the processor to direct the base station to perform the method according to any one of Examples 12 to 18.

[0150] Example 22: A computer-readable medium including instructions that, when executed by a processor, cause: a network entity including the processor to perform the method according to any one of Examples 1 to 11; a user equipment including the processor to perform the method according to any one of Examples 12 to 18; and / or a base station including the processor to perform the method according to any one of Examples 12 to 18.

Claims

1. A method for adaptive end-to-end E2E machine learning (ML) configuration performed by a network entity, the E2E ML configuration forming an E2E deep neural network (DNN) for processing communication transmitted via E2E communication between at least two endpoints, the E2E communication using a wireless network, the method comprising: The user equipment (UE) participating in the E2E communication is guided to form at least a first part of the E2E DNN based on a first E2E ML configuration to achieve the E2E communication; The base station participating in the E2E communication is guided to implement the E2E communication by forming at least a second part of the E2E DNN based on the first E2E ML configuration; The first E2EML configuration is updated based on the change in the participation mode of the edge computing server (ECS) in the E2E communication. The second E2E ML configuration is identified based on the change in the participation mode of the ECS in the E2E communication; as well as The system guides at least the UE or the base station to use the second E2E ML configuration to update at least a third portion of the E2E DNN for implementing the E2E communication.

2. The method according to claim 1, wherein, Determining to update the first E2E ML configuration includes: Determine whether to include the ECS in the E2E communication; and The first E2E ML configuration is updated based on the determination to include the ECS in the E2E communication.

3. The method according to claim 2, wherein, Determining to update the first E2EML configuration based on determining that it includes the ECS further includes: The decision to update the first E2E ML configuration is based on the following: In the E2E communication, communication with the ECS and communication with remote services are aggregated; or Exclude the remote service from the E2E communication.

4. The method according to claim 3, wherein, Determining to update the first E2E ML configuration includes: determining to update the first E2E ML configuration based on the aggregation, and The second E2E ML configuration is identified by: The downlink E2E ML configuration that forms the downlink E2E DNN is identified as at least a part of the second E2E ML configuration, and the downlink E2E DNN is directed to: Receive the first portion of application data from the ECS; Receive a second portion of the application data from the remote service; and The first part and the second part are aggregated to generate aggregated application data that is directed to the UE.

5. The method according to claim 3, wherein, Determining to update the first E2E ML configuration includes: determining to update the first E2E ML configuration based on the aggregation, and The second E2E ML configuration is identified by: The uplink E2E ML configuration that forms the uplink E2E DNN is identified as at least a portion of the second E2E ML configuration, and the uplink E2E DNN is configured to: Receive uplink application data from the UE; The uplink application data is used to generate a first output that is directed to the ECS; and The uplink application data is used to generate a second output that is directed to the remote service.

6. The method according to claim 1, wherein, Determining to update the first E2E ML configuration further includes: Receive a request from the UE to include the ECS in the E2E communication; or The inclusion of the ECS in the E2E communication is determined based on the estimated UE location.

7. The method according to claim 1, wherein, The third part of guiding at least the UE or the base station to update the E2E DNN using the second E2E ML configuration further includes: The UE is instructed to use the second E2E ML configuration to update the first part of the E2E DNN; or The base station is instructed to use the second E2E ML configuration to update the second part of the E2E DNN.

8. The method according to any one of claims 1-7, wherein, Identifying the second E2E ML configuration includes at least one of the following: Identify changes to one or more parameters of the E2E DNN; and Identify one or more architectural changes to the E2E DNN.

9. A method for adaptive end-to-end E2E machine learning (ML) configuration performed by a wireless transmit / receive unit (WTRU) for processing communications transmitted via E2E communication in a wireless network, the method comprising: The deep neural network DNN is formed using at least a first portion of the first E2E ML configuration of the E2E DNN used to implement E2E communication; Based on the change in the participation mode of the edge computing server ECS in the E2E communication, an instruction is received to update the DNN using at least a second part of the second E2E ML configuration. The DNN is updated using at least a second portion of the second E2E ML configuration; as well as At least a portion of the E2E communication is implemented using an updated DNN.

10. The method of claim 9, further comprising: The edge computing server is identified based on the estimated UE location of the WTRU; as well as The request is to include the ECS in the E2E communication.

11. The method according to claim 9 or claim 10, wherein, Receiving the instruction to update the DNN further includes: Receive guidance to update one or more parameters of the DNN as the instruction; or The instruction is to receive guidance on updating the architecture of the DNN.

12. The method according to claim 11, wherein, Receiving the updated architecture further includes at least one of the following: Change the number of processing layers used in the DNN; and Change the computation mode of at least one processing layer in the DNN.

13. A network entity, comprising: processor; as well as A computer-readable storage medium comprising instructions for implementing an end-to-end machine learning controller for performing the method according to any one of claims 1 to 8.

14. A user equipment, comprising: processor; as well as A computer-readable storage medium including instructions that, when executed by the processor, cause the user equipment to perform the method according to any one of claims 9 to 12.

15. A base station, comprising: processor; as well as A computer-readable storage medium including instructions that, when executed by the processor, cause the base station to perform the method according to any one of claims 9 to 12.

Images