Operation related to AI / ML models

Abstract

Exemplary embodiments of the present disclosure relate to operations associated with artificial intelligence / machine learning (AI / ML) models. In one example method, a first network device sends a first request to a second network device, the first request instructing the second network device to provide a first AI / ML model, and receives the first AI / ML model from the second network device. The first network device then obtains a trimmed AI / ML model based on the first AI / ML model. In this manner, a custom local model may be obtained from a global base model to reduce training complexity at a random access network (RAN) node. At the radio access network node, a relatively lightweight custom AI / ML model may be obtained from a substantial global base model at a core network (CN) or a third party, thereby reducing the training complexity at the RAN node.

Description

Technical Field

[0001] The exemplary embodiments disclosed herein generally relate to the field of communications, and more specifically to operations related to artificial intelligence / machine learning (AI / ML) models. Background Technology

[0002] Artificial intelligence (AI), and especially deep machine learning (ML), is a broad branch of computer science that involves building intelligent machines capable of performing tasks that typically require human intelligence. The introduction of AI is expected to bring about a paradigm shift in almost every area of ​​the technology industry, with AI poised to play a role in advancements in network technologies. For example, existing communication technologies rely on classical analytical modeling of channels, enabling wireless communication to operate close to the theoretical Shannon limit. Existing technologies may not be satisfactory for further maximizing the efficient use of the signal space. AI is expected to help address this challenge. Other aspects of wireless communication may benefit from the use of AI, particularly in future generations of wireless technologies, such as advanced 5G and future 6G systems and beyond.

[0003] To support the use of AI in wireless networks, a suitable network architecture is required. Therefore, providing a network architecture that supports the use of AI in wireless communications would be extremely useful, both for current and future generations of wireless systems. With the increasing number of AI tasks in future networks, fragmented models would be too costly (as they require separate hardware) and inefficient if radio access network (RAN) nodes (e.g., BS) trained their own models for each AI task. Summary of the Invention

[0004] In general, the exemplary embodiments of this disclosure provide solutions for operations associated with artificial intelligence / machine learning (AI / ML) models, such as those for operations from core network (CN) nodes or third-party (3) rdA solution for a global base model at a multi-access edge computing (MEC) platform and a custom local artificial intelligence / machine learning (AI / ML) model at a random access network (RAN) node.

[0005] In a first aspect, a method is provided. The method includes: sending a first request from a first network device to a second network device, the first request instructing the second network device to provide a first artificial intelligence / machine learning (AI / ML) model; receiving the first AI / ML model from the second network device; and obtaining a fine-tuned AI / ML model based on the first AI / ML model. In this way, a relatively lightweight custom local AI / ML model can be obtained from a fairly large (and “bulky”) global base model at a CN node or a third party, thereby reducing the training complexity at the first network device.

[0006] In some example embodiments, obtaining a fine-tuned AI / ML model includes fine-tuning a first AI / ML model based on data from a first network device. This allows for the acquisition of a more accurate local AI / ML model.

[0007] In some example embodiments, the method further includes: sending a second request to a second network device, the second request instructing the second network device to provide a second AI / ML model; and receiving the second AI / ML model from the second network device. In this way, more than one task-specific AI / ML model can be obtained from the second network device.

[0008] In some example implementations, the second request is sent together with the first request, and the second AI / ML model is received together with the first AI / ML model. This reduces signaling overhead compared to sending the two requests separately.

[0009] In some example embodiments, receiving the second AI / ML model together with the first AI / ML model includes receiving: at least one model parameter common to both the first and second AI / ML models; at least one model parameter specific to the first AI / ML model; and at least one model parameter specific to the second AI / ML model. By receiving the common model parameter once instead of twice, i.e., separately for the first and second AI / ML models, signaling overhead can be reduced.

[0010] In some example embodiments, the method further includes: sending a fine-tuned AI / ML model to a second network device; and receiving an updated AI / ML model from the second network device. In this way, the AI / ML model at the first network device can perform a specific task more accurately.

[0011] In some exemplary embodiments, the method further includes: sending a fine-tuned AI / ML model to at least one terminal device; receiving at least one third AI / ML model from at least one terminal device; and generating an updated AI / ML model based on the at least one third AI / ML model. In this way, the AI / ML model at the first network device can perform a specific task more accurately, especially when the task is related to at least one terminal device.

[0012] In some example embodiments, at least one terminal device includes multiple terminal devices, at least one third AI / ML model includes multiple AI / ML models, and generating an updated AI / ML model based on at least one third AI / ML model includes aggregating multiple AI / ML models to generate a fourth AI / ML model as the updated AI / ML model. In this way, the AI / ML model at the first network device can perform specific tasks more accurately, especially when the task is related to at least one terminal device.

[0013] In some example embodiments, generating an updated AI / ML model based on at least one AI / ML model further includes aggregating a fourth AI / ML model and a fine-tuned AI / ML model to generate a fifth AI / ML model as the updated AI / ML model. In this way, the AI / ML model at the first network device can perform specific tasks more accurately, especially when the task is related to at least one terminal device.

[0014] In some example embodiments, the method further includes: sending an updated AI / ML model to a second network device; and receiving a further updated AI / ML model from the second network device. In this way, the AI / ML model at the first network device can perform a specific task more accurately.

[0015] In some example embodiments, the method further includes sending data to the second network device before receiving the AI / ML model from the second network device, wherein the AI / ML model received by the first network device is a fine-tuned AI / ML model based on the data. In this way, the AI / ML model obtained from the second network device is more accurate for the first network device.

[0016] In some example embodiments, the method further includes at least one of: performing a task using at least one of an AI / ML model, a fine-tuned AI / ML model, an updated AI / ML model, or a further updated AI / ML model; or sending data to a second network device, the data being associated with at least one of a plurality of tasks and stored in an AI / ML database at the first network device. In this way, the first network device can perform localized tasks using a more accurate AI / ML model.

[0017] In this way, according to the first aspect and its exemplary embodiments, a relatively lightweight custom local AI / ML model can be obtained from the rather large (and “bulky”) global base model at the second network device, thereby reducing the training complexity at the first network device. Simultaneously, the local AI / ML model at the first network device is more accurate, thus enabling the first network device to perform tasks more accurately.

[0018] In a second aspect, a method is provided. This method includes: receiving, at a second network device, a first request from a first network device instructing the second network device to provide an artificial intelligence / machine learning (AI / ML) model; generating an AI / ML model at the second network device based on a pre-trained AI / ML model, wherein the pre-trained AI / ML model is pre-trained for multiple tasks, including tasks to be performed by the first network device using the AI / ML model; and sending the AI / ML model to the first network device. In this way, the second network device does not need to send a relatively large (and “bulky” global base model to the first network device; instead, the second network device can send a relatively lightweight custom AI / ML model to the first network device. Therefore, the training complexity at the first network device can be significantly reduced. Simultaneously, the AI / ML model at the first network device is more accurate and “tailored” to the first network device, enabling the first network device to perform tasks more accurately.

[0019] In some example embodiments, the method further includes: receiving a second request from a first network device, the second request instructing a second network device to provide a second AI / ML model; generating the second AI / ML model based on a pre-trained AI / ML model; and sending the second AI / ML model to the first network device. In this way, the second network device can send more than one task-specific AI / ML model to the first network device.

[0020] In some example implementations, the second request is received together with the first request, and the second AI / ML model is sent together with the first AI / ML model. This reduces signaling overhead compared to sending the two requests separately.

[0021] In some example embodiments, sending the second AI / ML model together with the first AI / ML model includes sending: at least one model parameter common to both the first and second AI / ML models; at least one model parameter specific to the first AI / ML model; and at least one model parameter specific to the second AI / ML model. By sending the common at least one model parameter once instead of receiving it twice, i.e., separately for the first and second AI / ML models, signaling overhead can be reduced.

[0022] In some example embodiments, the method further includes: receiving at least one AI / ML model from at least one network device including a first network device, the at least one AI / ML model including an AI / ML model provided by the first network device; generating an updated AI / ML model based on the at least one AI / ML model; and sending the updated AI / ML model to the first network device. In this way, the AI / ML model at the first network device can perform a specific task more accurately.

[0023] In some example embodiments, the method further includes: receiving at least one AI / ML model from at least one network device including a first network device, the at least one AI / ML model including an updated AI / ML model provided by the first network device, wherein the updated AI / ML model is generated based on at least one AI / ML model provided by at least one terminal device; generating an updated AI / ML model based on the at least one AI / ML model; and sending the updated AI / ML model to the first network device. In this way, the AI / ML model at the first network device can perform a specific task more accurately, especially when the task is related to at least one terminal device.

[0024] In some example embodiments, the AI / ML model sent to the first network device is a fine-tuned AI / ML model based on data from the first network device. In this way, the AI / ML model at the first network device can perform specific tasks more accurately.

[0025] In some example embodiments, the method further includes: receiving data from the first network device for fine-tuning the AI / ML model before sending the AI / ML model to the first network device; and performing fine-tuning on the AI / ML model based on the received data to obtain a fine-tuned AI / ML pattern. In this way, the AI / ML model at the first network device can be more accurate and "tuned" to perform a specific task.

[0026] In some example embodiments, the method further includes: receiving data from a first network device, the data being related to at least one of a plurality of tasks and stored at the first network device; and storing the received data at a second network device. In this way, a global base model at the second network device can be trained using the received data to be more accurate for multiple tasks, and the second network device can generate a more custom AI / ML model specifically for the first network device.

[0027] In this way, according to the second aspect and its exemplary embodiments, a relatively lightweight custom AI / ML model can be provided to the first network device instead of a rather large (and "bulky") global base model, thereby reducing the training complexity at the first network device. Simultaneously, the AI / ML model at the first network device is more accurate, thus enabling the first network device to perform tasks more accurately. Furthermore, the second network device can use data received from the first network device to train the global base model for greater accuracy across multiple tasks.

[0028] In a third aspect, a method is provided. This method includes: receiving an artificial intelligence / machine learning (AI / ML) model from a first network device at a terminal device; fine-tuning the AI / ML model based on data collected at the terminal device to obtain an updated, fine-tuned AI / ML model; and sending the updated AI / ML model to the first network device. In this manner, the AI / ML model to be used by the first network device can perform tasks more accurately, particularly tasks related to the terminal device.

[0029] In a fourth aspect, a first network device is provided. The first network device includes: a transceiver; and a processor communicatively coupled to the transceiver, wherein the processor is configured to: send a first request to a second network device via the transceiver, the first request instructing the second network device to provide a first artificial intelligence / machine learning (AI / ML) model; receive the first AI / ML model from the second network device via the transceiver; and obtain a fine-tuned AI / ML model based on the first AI / ML model. In this way, a relatively lightweight custom local AI / ML model can be obtained from a rather large (and “bulky”) global base model at a CN node or a third party, thereby reducing the training complexity at the first network device.

[0030] In a fifth aspect, a second network device is provided. The second network device includes: a transceiver; and a processor communicatively coupled to the transceiver, wherein the processor is configured to: receive a first request from a first network device via the transceiver, the first request instructing the second network device to provide an artificial intelligence / machine learning (AI / ML) model; generate an AI / ML model based on a pre-trained AI / ML model at the second network device, wherein the pre-trained AI / ML model is pre-trained for multiple tasks, including tasks to be performed by the first network device using the AI / ML model; and transmit the AI / ML model to the first network device via the transceiver. In this way, the second network device does not need to send a considerably large (and “bulky” global base model to the first network device; instead, the second network device can send a relatively lightweight custom AI / ML model to the first network device. Therefore, the training complexity at the first network device can be significantly reduced. Simultaneously, the AI / ML model at the first network device is more accurate and “tailored” to the first network device, enabling the first network device to perform tasks more accurately.

[0031] In a sixth aspect, a terminal device is provided. The terminal device includes: a transceiver; and a processor communicatively coupled to the transceiver, wherein the processor is configured to: receive an artificial intelligence / machine learning (AI / ML) model from a first network device via the transceiver; fine-tune the AI / ML model based on data collected at the terminal device to obtain an updated AI / ML model; and transmit the updated AI / ML model to the first network device via the transceiver. In this manner, the AI / ML model to be used by the first network device can perform tasks more accurately, particularly tasks related to the terminal device.

[0032] In a seventh aspect, a non-transient computer-readable storage medium is provided, comprising a computer program stored thereon. When executed on at least one processor, the computer program causes the at least one processor to perform a method of any one of the first, second, or third aspects. In this way, instead of sending a considerably large (and “bulky”) global base model to the first network device, the second network device can send a relatively lightweight custom AI / ML model to the first network device. Therefore, the training complexity at the first network device can be significantly reduced. Simultaneously, the AI / ML model at the first network device is more accurate and “tailored” to the first network device, enabling the first network device to perform tasks more accurately.

[0033] In an eighth aspect, a chip including at least one processing circuit is provided for performing the method of any one of the first, second, or third aspects. In this way, instead of sending a considerably large (and "bulky") global base model to the first network device, the second network device can send a relatively lightweight custom AI / ML model to the first network device. Therefore, the training complexity at the first network device can be significantly reduced. Simultaneously, the AI / ML model at the first network device is more accurate and "tailored" to the first network device, enabling the first network device to perform tasks more accurately.

[0034] In a ninth aspect, a computer program product tangibly stored on a computer-readable medium and including computer-executable instructions is provided, which, when executed, cause a device to perform any one of the first, second, or third aspects. In this way, instead of sending a considerably large (and “bulky”) global base model to the first network device, the second network device can send a relatively lightweight custom AI / ML model to the first network device. Therefore, the training complexity at the first network device can be significantly reduced. Simultaneously, the AI / ML model at the first network device is more accurate and “tailored” to the first network device, enabling the first network device to perform tasks more accurately.

[0035] It should be understood that the summary portion is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0036] Some exemplary embodiments will now be described with reference to the accompanying drawings, in which: FIG. 1A Examples of network environments in which some exemplary embodiments of this disclosure can be implemented are shown; FIG. 1B An exemplary communication system 100B is shown that can implement some exemplary embodiments of the present disclosure; FIG. 1C Examples of electronic devices and base stations according to some exemplary embodiments of the present disclosure are shown; FIG. 1D This illustration shows units or modules in a device according to some exemplary embodiments of the present disclosure; FIG. 1E A wireless system implementing an exemplary network architecture according to some exemplary embodiments of the present disclosure is shown; FIG. 1F Another exemplary wireless system according to some exemplary embodiments of the present disclosure is shown; FIG. 1G Another exemplary wireless system according to some exemplary embodiments of the present disclosure is shown; FIG. 1H Example apparatuses that can implement the methods and teachings according to some exemplary embodiments of the present disclosure are shown; FIG. 1I A schematic diagram of an exemplary pre-trained large model is shown according to some exemplary embodiments of the present disclosure; FIG. 1J A simplified block diagram of an exemplary data flow in exemplary operation of an AI module according to some exemplary embodiments of the present disclosure is shown; FIG. 2 A signaling diagram illustrating an exemplary communication process according to some exemplary embodiments of the present disclosure is shown; FIG. 3 A schematic diagram illustrating exemplary AI model implementations according to some embodiments of the present disclosure is shown; FIG. 4 A signaling diagram illustrating another exemplary communication process according to some embodiments of the present disclosure is shown; FIG. 5 A signaling diagram illustrating another exemplary communication process according to some embodiments of the present disclosure is shown; FIG. 6 A signaling diagram illustrating another exemplary communication process according to some embodiments of the present disclosure is shown; FIG. 7 A signaling diagram illustrating another exemplary communication process according to some embodiments of the present disclosure is shown; FIG. 8 A signaling diagram illustrating another exemplary communication process is shown, according to some exemplary embodiments of the present disclosure; FIG. 9 A flowchart illustrating an exemplary method implemented at a first network device according to some embodiments of the present disclosure is shown; FIG. 10 Another flowchart illustrates an exemplary method implemented at a second network device according to some embodiments of the present disclosure; FIG. 11 Another flowchart of an exemplary method implemented at a terminal device according to some embodiments of the present disclosure is shown; FIG. 12 A simplified block diagram of an apparatus according to some exemplary embodiments of the present disclosure is shown; FIG. 13 A simplified block diagram of another apparatus according to some exemplary embodiments of the present disclosure is shown; FIG. 14 A simplified block diagram of another apparatus according to some exemplary embodiments of the present disclosure is shown; and FIG. 15 A simplified block diagram of an apparatus suitable for implementing some exemplary embodiments of the present disclosure is shown.

[0037] Throughout the accompanying drawings, the same or similar reference numerals denote the same or similar elements. Detailed Implementation

[0038] The principles of this disclosure will now be described with reference to some exemplary embodiments. It should be understood that these embodiments are described merely for illustration and to help those skilled in the art understand and implement this disclosure, and do not impose any limitation on the scope of this disclosure. The disclosure described herein can be implemented in a variety of ways other than those described below.

[0039] In the following description and claims, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains.

[0040] References to "an embodiment," "embodiment," "exemplary embodiment," etc., in this disclosure indicate that the described embodiments may include specific features, structures, or characteristics, but not every embodiment must include specific features, structures, or characteristics. Furthermore, these phrases do not necessarily refer to the same embodiment. Moreover, when a specific feature, structure, or characteristic is described in connection with an embodiment, it should be understood that, whether explicitly described or not, those skilled in the art will recognize how such a feature, structure, or characteristic can be combined with other embodiments to achieve the desired result.

[0041] It should be understood that while the terms “first” and “second” may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used to distinguish one element from another. For example, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element, without departing from the scope of the exemplary embodiments. The term “and / or” as used herein includes any and all combinations of one or more of the listed items.

[0042] The terminology used herein is for describing particular embodiments and is not intended to limit the exemplary embodiments. Unless the context clearly indicates otherwise, the singular forms “a,” “an,” and “the” used herein are intended to include the plural forms as well. It should also be understood that the terms “comprises / comprising,” “has / having,” and / or “includes / including”, when used herein, specify the presence of the stated features, elements, and / or components, but do not exclude the presence or addition of one or more other features, elements, components, and / or combinations thereof.

[0043] As used herein, the term "communication network" refers to a network that conforms to any suitable communication standard, such as Long Term Evolution (LTE), LTE-Advanced (LTE-A), Wideband Code Division Multiple Access (WCDMA), High-Speed ​​Packet Access (HSPA), Narrow Band Internet of Things (NB-IoT), Wireless Fidelity (WiFi), etc. Furthermore, communication between terminal devices and network devices within a communication network can be performed according to any suitable generation communication protocol, including but not limited to fourth-generation (4G), 4.5G, the future fifth-generation (5G) IEEE 802.11 communication protocol, and / or any other currently known or future protocols. Embodiments of this disclosure can be applied to various communication systems. Given the rapid development of communication technologies, new communication technologies and systems embodying this disclosure will inevitably emerge in the future. The scope of this disclosure should not be limited to the systems described above.

[0044] As used herein, the term "network device" refers to a node in a communication network through which terminal devices access the network and receive services. Network devices can refer to base stations (BS) or access points (APs), such as node B (NodeB or NB), evolved NodeB (eNodeB or eNB), NR NB (also known as gNB), Remote Radio Unit (RRU), radio header (RH), remote radio head (RRH), WiFi device, repeater, low-power node (such as femtonode, piconode), etc., depending on the terminology and technology applied. In the following description, the terms "network device," "AP device," "AP," and "access point" are used interchangeably.

[0045] The term "terminal equipment" refers to any terminal device capable of wireless communication. By way of example and not limitation, terminal equipment may also be referred to as communication equipment, user equipment (UE), subscriber station (SS), portable subscriber station, mobile station (MS), station (STA), or station equipment or access terminal (AT). Terminal devices may include, but are not limited to, mobile phones, cellular phones, smartphones, Voice over IP (VoIP) phones, wireless local loop phones, tablets, wearable terminal devices, personal digital assistants (PDAs), portable computers, desktop computers, image capture terminal devices (e.g., digital cameras), gaming terminal devices, music storage and playback devices, in-vehicle wireless terminal devices, wireless endpoints, mobile stations, laptop-embedded equipment (LEE), laptop-mounted equipment (LME), USB dongles, smart devices, customer-premises equipment (CPE), Internet of Things (IoT) devices, watches or other wearable devices, virtual reality (VR) devices, extended reality (XR) devices, and head-mounted displays. Display (HMD), vehicles, drones, medical devices and applications (e.g., remote surgery), industrial devices and applications (e.g., robots and / or other wireless devices operating in the context of industrial and / or automated processing chains), consumer electronic devices, devices operating on commercial and / or industrial wireless networks, etc. In the following description, the terms “site,” “site equipment,” “STA,” “terminal equipment,” “communication equipment,” “terminal,” “user equipment,” and “UE” are used interchangeably.

[0046] refer to FIG. 1AThis diagram, provided as an illustrative example and not as limiting, is a simplified schematic of a communication system. Communication system 100A includes a radio access network 120. Radio access network 120 may be a next-generation (e.g., sixth-generation (6G) or later) radio access network, or a traditional (e.g., 5G, 4G, 3G, or 2G) radio access network. One or more user equipment (UE, also known as electronic device (ED)) 110a to 120j (generally referred to as 110) may interconnect with each other or be connected to one or more network nodes (170a, 170b, generally referred to as 170) within radio access network 120. Core network 130 may be part of the communication system and may depend on or be independent of the radio access technology used in communication system 100. Furthermore, communication system 100 includes a public switched telephone network (PSTN) 180, the Internet 150, and other networks 160. Other networks 160 may include a multi-access edge computing (MEC) platform, which will be described in more detail later.

[0047] FIG. 1B An exemplary communication system 100B is illustrated. Generally, the communication system 100 enables multiple wireless or wired components to transmit data and other content. The purpose of the communication system 100 may be to provide content such as voice, data, video, and / or text via broadcast, multicast, and unicast. The communication system 100 can operate by sharing resources (e.g., carrier spectrum bandwidth) among its constituent units. The communication system 100 may include terrestrial communication systems and / or non-terrestrial communication systems. The communication system 100 can provide a wide range of communication services and applications (e.g., earth monitoring, remote sensing, passive sensing and positioning, navigation and tracking, automated delivery and mobility, etc.). The communication system 100 can provide high availability and robustness through the joint operation of terrestrial and non-terrestrial communication systems. For example, integrating a non-terrestrial communication system (or components thereof) into a terrestrial communication system can create a heterogeneous network that can be considered as comprising multiple layers. Compared to traditional communication networks, heterogeneous networks can achieve better overall performance through efficient multi-link joint operation, more flexible function sharing, and faster physical layer link switching between terrestrial and non-terrestrial networks.

[0048] Terrestrial and non-terrestrial communication systems can be considered as subsystems of a communication system. In the example shown, communication system 100 includes electronic devices (EDs) 110a to 110d (generally referred to as ED 110), radio access networks (RANs) 120a to 120b, a non-terrestrial communication network 120c, a core network 130, a public switched telephone network (PSTN) 180, the Internet 150, and other networks 160. RANs 120a to 120b include corresponding base stations (BSs) 170a to 170b, which can generally be referred to as terrestrial transmit and receive points (T-TRPs) 170a to 170b. The non-terrestrial communication network 120c includes access nodes 120c, which can generally be referred to as non-terrestrial transmit and receive points (NT-TRPs) 172. As mentioned above, other networks 160 may include multi-access edge computing (MEC) platforms.

[0049] Alternatively or additionally, any ED 110 can be used to connect to, access, or communicate with any other T-TRP 170a to 170b, NT-TRP 172, Internet 150, core network 130, PSTN 180, other network 160, or any combination thereof. In some examples, ED 110a can communicate uplink and / or downlink with T-TRP 170a via interface 190a. In some examples, ED 110a, ED 110b, and ED 110d can also communicate directly with each other via one or more side-channel air interfaces 190b. In some examples, ED 110d can communicate uplink and / or downlink with NT-TRP 172 via interface 190c.

[0050] Air interfaces 190a and 190b can use similar communication technologies, such as any applicable wireless access technology. For example, communication system 100 can implement one or more channel access methods in air interfaces 190a and 190b, such as code division multiple access (CDMA), time division multiple access (TDMA), frequency division multiple access (FDMA), orthogonal FDMA (OFDMA), or single-carrier FDMA (SC-FDMA). Air interfaces 190a and 190b can utilize other higher-dimensional signal spaces, which may involve combinations of orthogonal and / or non-orthogonal dimensions.

[0051] The 190c air interface enables communication between the ED 110d and one or more NT-TRP172s via a wireless link or simply via a link. For some examples, the link is a dedicated connection for unicast transmission, a connection for broadcast transmission, or a connection between a group of EDs and one or more NT-TRPs for multicast transmission.

[0052] RAN 120a and RAN 120b communicate with core network 130 to provide various services, such as voice, data, and other services, to ED 110a, ED 110b, and ED 110c. RAN 120a and RAN 120b and / or core network 130 may communicate directly or indirectly with one or more other RANs (not shown), which may or may not be directly served by core network 130, and may or may not use the same radio access technology as RAN 120a, RAN 120b, or both. Core network 130 may also serve as a gateway access between (i) RAN 120a and RAN 120b or ED 110a, ED 110b, and ED 110c, or both, and (ii) other networks (e.g., PSTN 180, Internet 150, and other networks 160). Additionally, some or all of ED 110a, ED 110b, and ED 110c may include the ability to communicate with different wireless networks via different wireless links using different wireless technologies and / or protocols. ED 110a, ED 110b, and ED 110c may communicate with a service provider or exchange (not shown) via a wired communication channel and with the Internet 185, rather than wirelessly (or also wirelessly). PSTN 140 may include a circuit-switched telephone network for providing plain old telephone service (POTS). The Internet 185 may include computer networks and / or subnets (internal networks) and include protocols such as Internet Protocol (IP), Transmission Control Protocol (TCP), and User Datagram Protocol (UDP). ED 110a, ED 110b, and ED 110c may be multimode devices capable of operating according to multiple wireless access technologies and include multiple transceivers required to support these technologies.

[0053] FIG. 1CAnother example of ED 110 and base stations 170a, 170b, and / or 170c is shown. ED 110 is used to connect people, objects, machines, etc. ED 110 can be widely used in various scenarios, such as cellular communication, device-to-device (D2D), vehicle-to-everything (V2X), peer-to-peer (P2P), machine-to-machine (M2M), machine-type communication (MTC), Internet of Things (IoT), virtual reality (VR), augmented reality (AR), industrial control, autonomous driving, telemedicine, smart grids, smart furniture, smart offices, smart wearables, smart transportation, smart cities, drones, robots, remote sensing, passive sensing, positioning, navigation and tracking, automated delivery and mobility, etc.

[0054] Each ED 110 represents any suitable end-user equipment used for wireless operation, which may include (or be referred to as) user equipment (UE / user device), wireless transmit / receive unit (WTRU), mobile station, fixed or mobile subscriber unit, cellular phone, station (STA), machine type communication (MTC) device, personal digital assistant (PDA), smartphone, laptop, computer, tablet, wireless sensor, consumer electronics, smart book, vehicle, automobile, truck, bus, train, or IoT device, industrial equipment or apparatus of the above (e.g., communication module, modem, or chip), etc. Future generations of ED 110 may be referred to using other terms. Base stations 170a and 170b are T-TRPs and will be referred to as T-TRP 170 below. Similarly, FIG. 3 As shown, NT-TRP is referred to below as NT-TRP 172. Each ED 110 connected to T-TRP 170 and / or NT-TRP 172 can be configured to be dynamically or semi-statically turned on (i.e., established, activated, or enabled), turned off (i.e., released, deactivated, or disabled), and / or in response to one or more of connection availability and connection necessity.

[0055] ED 110 includes a transmitter 201 and a receiver 203 coupled to one or more antennas 204. Only one antenna 204 is shown in the figure. One, part, or all of the antennas may also be panels. The transmitter 201 and receiver 203 may, for example, be integrated as a transceiver. The transceiver is used to modulate data or other content for transmission through at least one antenna 204 or a network interface controller (NIC). The transceiver is also used to demodulate data or other content received through at least one antenna 204. Each transceiver includes any suitable structure for generating signals for wireless or wired transmission and / or for processing signals received wirelessly or wiredly. Each antenna 204 includes any suitable structure for transmitting and / or receiving wireless or wired signals.

[0056] ED 110 includes at least one memory 208. Memory 208 stores instructions and data used, generated, or acquired by ED 110. For example, memory 208 may store software instructions or modules for implementing some or all of the functions and / or embodiments described herein, and executed by one or more processing units 210. Each memory 208 includes any suitable one or more volatile and / or non-volatile storage and retrieval devices. Any suitable type of memory can be used, such as random access memory (RAM), read-only memory (ROM), hard disk, optical disk, subscriber identity module (SIM) card, memory stick, secure digital (SD) memory card, and processor cache, etc.

[0057] ED 110 may also include one or more input / output devices (not shown) or interfaces (e.g., connected to...). FIG. 1A (Wired interface of Internet 185 in the network). Input / output devices support interaction with users or other devices on the network. Each input / output device includes any suitable structure for providing or receiving information from the user, such as a speaker, microphone, keypad, keyboard, display, or touchscreen, including network interface communication.

[0058] ED 110 also includes a processor 210 for performing various operations, including operations related to preparing for uplink transmissions to NT-TRP 172 and / or T-TRP 170, operations related to processing downlink transmissions received from NT-TRP 172 and / or T-TRP 170, and operations related to processing sidelink transmissions to and from another ED 110. Processing operations related to preparing for uplink transmissions may include operations such as encoding, modulation, transmit beamforming, and generating symbols for transmission. Processing operations related to processing downlink transmissions may include operations such as receive beamforming, demodulation, and decoding of received symbols. According to an embodiment, receiver 203 may receive downlink transmissions (possibly using receive beamforming), and processor 210 may extract signaling from the downlink transmissions (e.g., by detecting and / or decoding signaling). Examples of signaling may be reference signals transmitted by NT-TRP 172 and / or T-TRP 170. In some embodiments, processor 276 performs transmit beamforming and / or receive beamforming based on beam direction indications (e.g., beam angle information (BAI)) received from T-TRP 170. In some embodiments, processor 210 may perform operations related to network access (e.g., initial access) and / or downlink synchronization, such as operations related to detecting synchronization sequences, decoding, and acquiring system information. In some embodiments, processor 210 may perform channel estimation, for example, using reference signals received from NT-TRP 172 and / or T-TRP 170.

[0059] Although not shown, processor 210 may form part of transmitter 201 and / or receiver 203. Although not shown, memory 208 may form part of processor 210.

[0060] The processor 210, as well as the processing components of the transmitter 201 and receiver 203, may be implemented by the same or different one or more processors, which execute instructions stored in memory (e.g., memory 208). Alternatively, some or all of the processor 210, as well as the processing components of the transmitter 201 and receiver 203, may be implemented using special-purpose circuitry such as a field-programmable gate array (FPGA), a graphics processing unit (GPU), or an application-specific integrated circuit (ASIC).

[0061] In some implementations, T-TRP 170 may be referred to by other names, such as base station, base transceiver station (BTS), wireless base station, network node, network equipment, network-side equipment, transmit / receive node, Node B, evolved NodeB (eNodeB or eNB), home eNodeB, next-generation NodeB (gNB), transmission point (TP), site controller, access point (AP) or wireless router, relay station, remote radio head, ground node, ground network equipment or ground base station, base band unit (BBU), remote radio unit (RRU), active antenna unit (AAU), remote radio head (RRH), central unit (CU), distributed unit (DU), location node, etc. T-TRP 170 can be a macro BS, micro BS, relay node, donor node, or a combination thereof. T-TRP 170 may refer to the aforementioned equipment or a device within the aforementioned equipment (e.g., a communication module, modem, or chip).

[0062] In some embodiments, the various parts of T-TRP 170 may be distributed. For example, some modules of T-TRP 170 may be located remotely from the device housing the antenna of T-TRP 170 and may be coupled to the device housing the antenna via a communication link (not shown), sometimes referred to as the fronthaul, such as the Common Public Radio Interface (CPRI). Therefore, in some embodiments, the term "T-TRP 170" may also refer to modules on the network side that perform processing operations such as ED 110 location determination, resource allocation (scheduling), message generation, and encoding / decoding; these modules are not necessarily part of the device housing the antenna of T-TRP 170. These modules may also be coupled to other T-TRPs. In some embodiments, T-TRP 170 may actually be multiple T-TRPs that work together, for example, through coordinated multicast transmissions, to serve ED 110.

[0063] T-TRP 170 includes at least one transmitter 252 and at least one receiver 254 coupled to one or more antennas 256. Only one antenna 256 is shown in the figure. One, part or all of the antennas may also be panels. The transmitter 252 and receiver 254 may be integrated as a transceiver. T-TRP 170 also includes a processor 260 for performing various operations, including operations related to: preparing a transmission for downlink transmission to ED 110, processing uplink transmissions received from ED 110, preparing a transmission for backhaul transmission to NT-TRP 172, and processing transmissions received from NT-TRP 172 via backhaul. Processing operations related to preparing a transmission for downlink or backhaul transmission may include operations such as encoding, modulation, precoding (e.g., MIMO precoding), transmit beamforming, and generating symbols for transmission. Processing operations related to processing received transmissions in the uplink or backhaul may include operations such as receive beamforming, demodulation, and decoding of received symbols. Processor 260 can also perform operations related to network access (e.g., initial access) and / or downlink synchronization, such as generating the contents of a synchronization signal block (SSB), generating system information, etc. In some embodiments, processor 260 also generates beam direction indications, such as BAI, which can be scheduled for transmission by scheduler 253. Processor 260 performs other network-side processing operations described herein, such as determining the location of ED 110, determining the location for deploying NT-TRP 172, etc. In some embodiments, processor 260 can generate signaling, for example, for configuring one or more parameters of ED 110 and / or one or more parameters of NT-TRP 172. Any signaling generated by processor 260 is transmitted by transmitter 252. Note that the term "signaling" as used herein may also be referred to as control signaling. Dynamic signaling can be transmitted in control channels, such as the physical downlink control channel (PDCCH). Static or semi-static higher-layer signaling can be included in packets transmitted in data channels such as the physical downlink shared channel (PDSCH).

[0064] Scheduler 253 may be coupled to processor 260. Scheduler 253 may be included within or operate separately from T-TRP 170, which may schedule uplink, downlink, and / or backlink transmissions, including issuing scheduling authorizations and / or configuring schedule-free (“configuration authorization”) resources. T-TRP 170 also includes memory 258 for storing information and data. Memory 258 stores instructions and data used, generated, or acquired by T-TRP 170. For example, memory 258 may store software instructions or modules executed by processor 260 for implementing some or all of the functions and / or embodiments described herein.

[0065] Although not shown, processor 260 may form part of transmitter 252 and / or receiver 254. Furthermore, processor 260 may implement scheduler 253, which is not shown in the figure. Although not shown, memory 258 may form part of processor 260.

[0066] The processor 260, scheduler 253, and the processing components of transmitter 252 and receiver 254 can each be implemented by the same or different one or more processors, which execute instructions stored in memory (e.g., memory 258). Alternatively, some or all of the processor 260, scheduler 253, and the processing components of transmitter 252 and receiver 254 can be implemented using dedicated circuitry such as FPGA, GPU, or ASIC.

[0067] Although the NT-TRP 172 is shown as a drone only as an example, the NT-TRP 172 can be implemented in any suitable non-terrestrial form. Furthermore, in some implementations, the NT-TRP 172 may have other names, such as a non-terrestrial node, a non-terrestrial network device, or a non-terrestrial base station. The NT-TRP 172 includes a transmitter 272 and a receiver 274 coupled to one or more antennas 280. Only one antenna 280 is shown in the figure. One, part, or all of the antennas may also be panels. The transmitter 272 and receiver 274 may be integrated as a transceiver. The NT-TRP 172 also includes a processor 276 for performing various operations, including operations related to: preparing transmissions for downlink transmissions to ED 110, processing uplink transmissions received from ED 110, preparing transmissions for backhaul transmissions to T-TRP 170, and processing transmissions received from T-TRP 170 via backhaul. Processing operations related to preparing for downlink or backhaul transmissions may include operations such as encoding, modulation, precoding (e.g., MIMO precoding), transmit beamforming, and generating symbols for transmission. Processing operations related to receiving transmissions in the uplink or backhaul may include operations such as receive beamforming, demodulation, and decoding of received symbols. In some embodiments, processor 276 performs transmit beamforming and / or receive beamforming based on beam direction information (e.g., BAI) received from T-TRP 170. In some embodiments, processor 276 may generate signaling, for example, to configure one or more parameters of ED 110. In some embodiments, NT-TRP 172 implements physical layer processing but does not implement higher-level functions such as medium access control (MAC) or radio link control (RLC) layer functions. Since this is only an example, NT-TRP 172 may generally implement higher-level functions in addition to physical layer processing.

[0068] The NT-TRP 172 also includes a memory 278 for storing information and data. Although not shown, a processor 276 may form part of the transmitter 272 and / or receiver 274. Although not shown, the memory 278 may form part of the processor 276.

[0069] The processor 276, and the processing components of the transmitter 272 and receiver 274, may each be implemented by the same or different one or more processors, which execute instructions stored in memory (e.g., memory 278). Alternatively, some or all of the processor 276, and the processing components of the transmitter 272 and receiver 274, may be implemented using dedicated circuitry such as a programmed FPGA, GPU, or ASIC. In some embodiments, the NT-TRP 172 may actually be, for example, multiple NT-TRPs working together to serve ED 110 by coordinating multipoint transmissions.

[0070] T-TRP 170, NT-TRP 172 and / or ED 110 may include other components, but these components have been omitted for clarity.

[0071] according to FIG. 1D One or more steps in the methods of the embodiments provided herein may be performed by the corresponding units or modules. FIG. 1D The diagram illustrates units or modules within a device, such as ED 110, T-TRP 170, or NT-TRP 172. For example, signals may be transmitted by a transmitting unit or transmitting module. Signals may be received by a receiving unit or receiving module. Signals may be processed by a processing unit or processing module. Other steps may be performed by an artificial intelligence (AI) module or a machine learning (ML) module. The corresponding units or modules may be implemented using hardware, one or more components or devices executing software, or a combination thereof. For example, one or more of these units or modules may be integrated circuits, such as a programmed FPGA, GPU, or ASIC. It should be understood that if these modules are implemented by a processor using software, these modules may be retrieved by the processor, wholly or partially, individually or collectively, for processing, in one or more instances, and these modules themselves may include instructions for further deployment and instantiation.

[0072] Further details regarding ED 110, T-TRP 170, and NT-TRP 172 are known to those skilled in the art. Therefore, these details are omitted herein.

[0073] FIG. 1EA wireless system 100E with an implementation example network architecture according to embodiments of the present disclosure is illustrated. The wireless system 100E enables multiple wireless or wired components to transmit data and other content. The wireless system 100E allows content (e.g., voice, data, video, text, etc.) to be transmitted between entities of the system 100E (e.g., via broadcast, narrowcast, peer-to-peer, etc.). The wireless system 100E can operate by sharing resources such as bandwidth. The wireless system 100E can be adapted for wireless communication using 5G technology and / or next-generation wireless technologies (e.g., 6G or next-generation). In some examples, the wireless system 100E can also be compatible with some legacy wireless technologies (e.g., 3G or 4G wireless technologies).

[0074] In the example shown, the wireless system 100E includes multiple user equipment (UE) units 110, multiple system nodes 120, and a core network 130. The core network 130 can connect to a multi-access edge computing (MEC) platform 140 and one or more external networks 150 (e.g., the public switched telephone network (PSTN), the Internet, other private networks, etc.). Although FIG. 1E A certain number of these components or elements are shown, but the wireless system 100E may include any reasonable number of these components or elements.

[0075] Each UE 110 can be independently any suitable terminal device for wireless operation and can include (or be referred to as) electronic devices such as: wireless transmit / receive units (WTRUs), customer premises equipment (CPEs), smart devices, Internet of Things (IoT) devices, wireless-enabled vehicles, mobile stations, fixed or mobile subscriber units, cellular phones, stations (STAs), machine-type communication (MTC) devices, personal digital assistants (PDAs), smartphones, laptops, computers, tablets, wireless / wired sensors, or consumer electronic devices. Future generations of UE 110 may be referred to using other terms. For example, UE 110 may often be referred to as an electronic device (ED).

[0076] System node 120 can be any node in an access network (AN) (also known as a radio access network (RAN)). For example, system node 120 can be a base station (BS) of an AN. Each system node 120 is used to wirelessly connect with one or more UEs in UE 110 to enable access to the corresponding AN. A given UE 110 can connect to a given system node 120 to enable access to the core network 130, another system node 120, the MEC platform 140, and / or an external network 150. For example, system node 120 may include one or more of a number of well-known devices, such as a base transceiver station (BTS), a radio base station, a Node-B (NodeB), an evolved NodeB (eNodeB), a home base station, a gNodeB (sometimes called a next-generation Node B), a transmission point (TP), a transmit and receive point (TRP), a site controller, an access point (AP), an AP with sensing capabilities, a dedicated sensing node, or a wireless router, etc. System node 120 may also be or include mobile nodes, such as drones, unmanned aerial vehicles (UAVs), network-enabled vehicles (e.g., autonomous or semi-autonomous vehicles), etc. System node 120 may also be or include non-terrestrial nodes, such as satellites. Future generation system node 120 may include other network-enabled nodes and may be referred to using other terms.

[0077] Core network 130 may include one or more core servers or server clusters. Core network 130 provides core functions 132, such as core access and mobility management functions (AMF), user plane functions (UPF), and awareness management / control functions. Access to core functions 132 can be provided to UE 110 through the corresponding system node 120. Core network 130 can also serve as a gateway access between (i) system node 120 or UE 110 or both, and (ii) external network 150 and / or MEC platform 140. Core network 130 may provide a convergence interface (not shown), which is a common interface for all access types (e.g., wireless or wired access types).

[0078] MEC platform 140 can be a distributed computing platform in which multiple MEC hosts (typically edge servers) provide distributed computing resources (e.g., memory and processor resources). MEC platform 140 can provide functions and services closer to the end user (e.g., physically closer to system node 120 compared to core network 130), which can help reduce latency in providing such functions and services.

[0079] FIG. 1E Network node 131 is also shown, which can be any node on the network side of wireless system 100A (i.e., not any node of UE 110). For example, network node 131 can be a node of MEC platform 140 (e.g., MEC host), a node of external network 150 (e.g., network server), or a node within core network 130 (e.g., core server), etc. Network node 131 can be located outside core network 130 but directly connected to core network 130. Network node 131 can be a node connecting core network 130 and system node 120 (e.g., outside but close to AN, or within one or more ANs). Network node 131 can be dedicated to supporting AI capabilities (e.g., dedicated to performing the AI ​​management functions disclosed herein) and can be, for example, a link between multiple entities of wireless system 100A (including external network 150 and MEC platform 140, although for brevity such a link is not shown). FIG. 1A Access is shown in the diagram. It should be noted that while this disclosure provides examples of network node 131 providing certain AI functions (e.g., AI management module 210, discussed further below), the functions of network node 131 or similar AI functions (e.g., functions more focused on execution and less focused on training) may be provided by system node 120 or UE 110. For example, functions described as being provided at network node 131 may be additionally or alternatively provided at system node 120 or UE 110 as integrated / embedded functions or dedicated AI functions. Furthermore, network node 131 may have its own sensing capabilities and / or dedicated sensing nodes (not shown) to obtain sensing information (e.g., network data) for AI operation. In some examples, network node 131 may be an AI-dedicated node capable of performing higher-intensity and / or large-scale computations (which may be required for full training of the AI ​​model). Moreover, although shown as a single network node 131, it should be understood that network node 131 may actually be a representation of a distributed computing system (i.e., network node 131 may actually be a group of multiple physical computing systems), and not necessarily a single physical computing system. It should also be understood that network node 131 may include future network nodes that can be used in next-generation wireless technologies.

[0080] System node 120 communicates with one or more corresponding UEs 110 via AN-UE interface 125 (typically an air interface, such as radio frequency (RF), microwave, infrared (IR), etc.). For example, the RAN-UE interface may be a Uu link (e.g., based on 5G or 4G wireless technology). UEs 110 may also communicate directly with each other via one or more sidelink interfaces (not shown). Each system node 120 communicates with the core network 130 via AN-core network (CN) interface 135 (e.g., an NG interface based on 5G technology). Network node 131 may communicate with the core network 130 via dedicated interface 145, which will be discussed further below. Communication between system node 120 and core network 130, between two (or more) system nodes 120, and / or between network node 131 and core network 130 may be performed via backhaul links. Communication from UE 110 to system node 120 to core network 130 can be called uplink (UL) communication, and communication from core network 130 to system node 120 to UE 110 can be called downlink (DL) communication.

[0081] FIG. 1E An exemplary disclosed architecture is shown that can implement the AI ​​management module 210 and the AI ​​execution module 220. Other exemplary architectures will now be discussed.

[0082] FIG. 1F A wireless system 100B implementing another exemplary network architecture according to embodiments of the present disclosure is shown. It should be understood that... FIG. 1F Network architecture and FIG. 1E The network architectures have many similarities and do not need to repeat the details of common components.

[0083] and FIG. 1E Compared to the example shown, FIG. 1FThe network architecture of the wireless system 100F enables network node 131 to directly interface with each system node 120 (e.g., with at least one system node 120 of each AN) via interface 147, at which the AI ​​management module 210 is implemented. Interface 147 can be a public API interface or a dedicated interface for AI-related communications (e.g., communications using AI-related protocols such as those disclosed herein). It should be noted that interface 147 enables direct communication between the AI ​​management module 210 and the AI ​​execution module 220 at each system node 120 (regardless of whether network node 131 is a node in the MEC platform 140 or an external network 150, or whether network node 131 is part of the core network 130). Interface 147 can be a wired or wireless interface, such as a backhaul link between network node 131 and system node 120. Interface 147 is typically not found in 4G or 5G wireless systems. FIG. 1F Network node 131 can also be accessed by external network 150, MEC platform 140, and / or core network 130 (although for simplicity, such links are not shown in the text). FIG. 1F (As shown in the image).

[0084] FIG. 1G A wireless system 100G implementing another exemplary network architecture according to embodiments of the present disclosure is shown. It should be understood that... FIG. 1G Network architecture and FIG. 1E and 1F The network architectures have many similarities and do not need to repeat the details of common components. FIG. 1G An exemplary architecture is shown in which the AI ​​management module 210 is located in a network node 131 that is physically close to one or more system nodes 120 of one or more ANs managed by the AI ​​management module 210. For example, the network node 131 may be located in the same location as or within the MEC platform 140, or it may be located in the same location as or within the AN.

[0085] and FIG. 1E and 1F Compared to the example shown, FIG. 1G The 100G wireless system network architecture omits the AI ​​execution module 220 from system node 120. Alternatively, one or more local AI models (and optional local AI databases) that would otherwise be maintained in the local memory of each system node 120 could be maintained instead in the local memory of network node 131 (e.g., in the memory of the MEC host, or in distributed memory on the MEC platform 140). Although in FIG. 1GAlthough not shown in the diagram, network node 131 may implement one or more AI execution modules 220, or may implement the functions of AI execution module 220 in addition to AI management module 210, such as to collect network data and train and execute AI models in near real-time, and / or to separate global and local AI models.

[0086] Because network node 131 is physically located close to system node 120, communication between each system node 120 (e.g., from one or more ANs) and network node 131 can occur with very low latency (e.g., on the order of a few microseconds or milliseconds). Therefore, communication between system node 120 and network node 131 can occur in near real-time. As described above, communication between each system node 120 and network node 131 can be via interface 147. Interface 147 can be an AI-specific communication interface that supports low-latency communication.

[0087] FIG. 1H Exemplary apparatuses for implementing the methods and teachings according to this disclosure are shown. Specifically, FIG. 1H An exemplary computing system 250 is shown, which can be used to implement UE 110, system node 120, or network node 131. As will be discussed further below, computing system 250 may be dedicated to, or include, specialized components to support the training and / or execution of AI models (e.g., training and / or execution of neural networks).

[0088] like FIG. 1HAs shown, the computing system 250 includes at least one processing unit 251. The processing unit 251 implements various processing operations of the computing system 250. For example, the processing unit 251 may perform signal encoding, data processing, power control, input / output processing, or any other function of the computing system 250. Furthermore, the processing unit 251 may also be used to perform the computations required to train and / or execute AI models. In some examples, the processing unit 251 may be a dedicated processing unit capable of performing a large number of computations for training AI models. The processing unit 251 may, for example, include a microprocessor, microcontroller, digital signal processor, field-programmable gate array, application-specific integrated circuit, neural processing unit (NPU), tensor processing unit (TPU), or graphics processing unit (GPU). In some examples, the computing system 250 may have multiple processing units 251, wherein at least one processing unit 251 is a central processing unit (CPU) responsible for performing the core functions of the computing system 250 (e.g., the execution of an operating system (OS)), and at least another processing unit 251 is responsible for performing special functions (e.g., performing computations for training and / or executing AI models).

[0089] The computing system 250 includes at least one communication interface 252 for wired and / or wireless communication. Each communication interface 252 includes any suitable structure for generating signals for wireless or wired transmission and / or for processing signals received wirelessly or wiredly. In this example, the computing system 250 includes at least one antenna 254, for example, for the wireless communication interface 252 (in other examples, the antenna 254 may be omitted, for example, for the wired communication interface 252). Each antenna 254 includes any suitable structure for transmitting and / or receiving wireless or wired signals. One or more communication interfaces 252 may be used in the computing system 250. One or more antennas 254 may be used in the computing system 250. In some examples, the one or more antennas 254 may be an antenna array that can be used to perform beamforming and beam control operations. Although shown as a single functional unit, the communication interface 252 may also be implemented using at least one transmitter interface and at least one separate receiver interface. A processing unit 251 is coupled to the communication interface 252, for example, to provide data to be transmitted and / or received through the communication interface 252. The processing unit 251 can also control the operation of the communication interface 252 (e.g., to set parameters for wireless signaling).

[0090] The computing system 250 may include one or more optional input / output devices 256. Input / output devices 256 may interact with a user and / or optionally directly with other nodes such as UE 110, system node 120 (e.g., a base station), network node 131, or functional nodes in the core network 130. Each input / output device 256 may include any suitable structure for providing or receiving information from the user, such as a speaker, microphone, keypad, keyboard, display, or touchscreen, etc. A processing unit 251 is coupled to the input / output devices 256, for example, to provide data to be output via an output device or to receive data input via an input device.

[0091] The computing system 250 includes at least one memory 258. The memory 258 stores instructions and data used, generated, and / or collected by the computing system 250. For example, the memory 258 may store software instructions or modules for implementing some or all of the functions and / or embodiments described herein. A processing unit 251 is coupled to the memory 258 to enable the processing unit 251 to execute instructions stored in the memory 258 and, for example, to store data in the memory 258. The memory 258 includes any suitable one or more volatile and / or non-volatile storage and retrieval devices. Any suitable type of memory can be used, such as random access memory (RAM), read-only memory (ROM), hard disk, optical disk, subscriber identity module (SIM) card, memory stick, secure digital (SD) memory card, etc.

[0092] Refer again FIG. 1A The AI ​​capabilities in the wireless system 100A are supported by functions provided by an AI management module 210 and at least one AI execution module 220. The AI ​​management module 210 and the AI ​​execution module 220 are software modules that can be encoded into instructions stored in memory and executed by a processing unit.

[0093] In the example shown, the AI ​​management module 210 resides in network node 131, which may be located in the same location as MEC 140 or within MEC 140 (e.g., implemented on an MEC host, or implemented in a distributed manner across multiple MEC hosts). In other examples, the AI ​​management module 210 may reside in network node 131 as a node of external network 150 (e.g., implemented in a web server of external network 150). Generally, the AI ​​management module 210 may reside in any suitable network node 131, and may be located in network node 131 as part of core network 130 or outside of core network 130. In some examples, positioning the AI ​​management module 210 in network node 131 outside of core network 130 may enable a more open interface with external network 150 and / or third-party services, although this is not required. The AI ​​management module 210 can manage a large number of different AI models designed for different tasks, as discussed further below. Although the AI ​​management module 210 is shown within a single network node 131, it should be understood that the AI ​​management module 210 can also be implemented in a distributed manner (e.g., distributed across multiple network nodes 131, or the network node 131 itself is a representation of a distributed computing system).

[0094] In this example, each system node 120 implements a corresponding AI execution module 220. For example, system node 120 can be a BS within an AN, and can implement AI execution module 220 and perform the functionality of AI execution module 220 on behalf of the entire AN (or on behalf of a part of the AN). In another example, each BS within an AN can be a system node 120 that implements its own AI execution module 220. Therefore, FIG. 1A The multiple system nodes 120 shown may belong to the same AN ​​or may not belong to the same AN. In another example, system node 120 may be a single AI-capable node in the AN (i.e., not a BS), which may or may not be dedicated to providing AI functionality. Although each AI execution module 220 is shown within a single system node 120, it should be understood that each AI execution module 220 may be implemented independently and optionally in a distributed manner (e.g., distributed across multiple system nodes 120, or the system node 120 itself may be a representation of a distributed computing system).

[0095] AI execution module 220 can interact with some or all of the software modules of system node 120. For example, AI execution module 220 can interface with logical layers, such as the physical (PHY) layer, media access control (MAC) layer, radio link control (RLC) layer, packet data convergence protocol (PDCP) layer, and / or upper layers of system node 120 (at system node 120, the logical layer can be functionally divided into a high-level centralized unit (CU) layer and a low-level distributed unit (DU) layer). For example, AI execution module 220 can use a common application programming interface (API) to interface with the control module of system node 120.

[0096] Alternatively, UE 110 may also implement its own AI execution module 220. The AI ​​execution module 220 implemented by UE 110 can perform functions similar to the AI ​​execution module 220 implemented at system node 120. Other implementations are also possible. It should be noted that different UEs 110 may have different AI capabilities. For example, all, some, one, or none of the UEs 110 in wireless system 100A may implement the corresponding AI execution module 220.

[0097] In this example, network node 131 can communicate with one or more system nodes 120 via core network 130 (e.g., using AMF or / and UPF provided by core functionality 132 of core network 130). Network node 131 may have a communication interface with core network 130 using interface 145, which may be a public API interface or a dedicated interface for AI-related communication (e.g., communication using AI-related protocols, such as those disclosed herein). It should be noted that interface 145 enables direct communication between network node 131 and core network 130 (regardless of whether network node 131 is inside, near, or outside core network 130) by bypassing a convergence interface (in this scenario, communication between core network 130 and all external networks 150 typically requires a convergence interface). In another embodiment, network node 131 is within core network 130, and interface 145 is a communication interface within core network 130, such as a public API interface. Interface 145 can be a wired or wireless interface, such as a backhaul link between network node 131 and core network 130. Interface 145 can be an interface that is not typically present in 4G or 5G wireless systems. Therefore, core network 130 can be used to forward or relay AI-related communication between AI execution module 220 at one or more system nodes 120 (and optionally, one or more UEs 110) and AI management module 210 at network node 131. In this way, AI management module 210 can be considered to provide a set of AI-related functions in parallel with the core functions 132 provided by core network 130.

[0098] AI-related communication between system node 120 and one or more UEs 110 can be conducted through existing interfaces, such as Uu links in 5G and 4G network systems, or through a dedicated AI air interface (e.g., using AI-related protocols on the AI-related logical layer, as described herein). For example, AI-related communication between system node 120 and UEs 110 served by system node 120 can be conducted through the dedicated AI air interface, while non-AI-related communication can be conducted through 5G or 4G Uu links.

[0099] FIG. 1I A schematic diagram of an exemplary pre-trained large model 100I according to some exemplary embodiments of this disclosure is shown. The pre-trained large model is also referred to as a global model or a base model. The pre-trained large model can be deployed on the core network (CN) or by a third party to support multiple tasks. Here, the pre-trained large model 100I is used as the basis for AI tasks on the radio access network (RAN) side.

[0100] like FIG. 1IAs shown, the pre-trained large model 100I is pre-trained for multiple tasks. When task-1 is input into the pre-trained large model, inference-1 corresponding to input task-1 is obtained. Similarly, when task-2 is input into the pre-trained large model, inference-2 corresponding to input task-2 is obtained. This continues. When task-N (where N is an integer greater than 2) is input into the pre-trained large model, inference-N corresponding to input task-N is obtained.

[0101] Currently, the number of AI tasks in future networks will increase. If RAN nodes (e.g., BS) train their own models for each AI task, fragmented models become too expensive (because each AI model requires separate hardware) and inefficient. In this scenario, the RAN can obtain a basic custom model from the global model (e.g., a smaller model than the global model) and perform fine-tuning on the local models. This is the basic technical concept disclosed herein, which will be referenced later. FIG. 2 to 15 To describe in more detail.

[0102] FIG. 1J It is shown, for example, in FIG. 1E and 1F The diagram shows a simplified block diagram of an exemplary data flow in the exemplary operation of the AI ​​management module 210 and the AI ​​execution module 220. In this example, the AI ​​execution module 220 is implemented in a system node 120, such as the BS of an AN. It should be understood that if the AI ​​execution module 220 is implemented in the UE 110, similar operations can be performed (and the system node 120 can be an intermediary relaying AI-related communication between the UE 110 and the network node 131). Furthermore, communication to and from the network node 131 may or may not be relayed through the core network 130.

[0103] The task request is received by the AI ​​management module 210. First, an example of a network task request is described. A network task request can be any request for a network task, including requests for services, and can include one or more task requirements, such as one or more KPIs (e.g., latency, QoS, throughput, etc.) and / or application attributes (e.g., traffic type, etc.) related to the network task. Task requests can be received from clients of the wireless system 100E or 100F, from the external network 150, and / or from nodes within the wireless system 100E or 100F (e.g., from system node 120 itself).

[0104] Upon receiving a task request, AI management module 210 performs functions (e.g., using functions provided by AIMF 212 and / or AICF 214) to perform initial setup and configuration based on the task request. For example, AI management module 210 can use the functions of AICF 214 to set target KPIs and application or traffic types for network tasks based on one or more task requirements included in the task request. Initial setup and configuration may include selecting one or more global AI models 216 (from multiple available global AI models 216 maintained by AI management module 210) to meet the task request. The global AI models 216 available to AI management module 210 may be developed, updated, configured, and / or trained by the operator of core network 130, other operators, external network 150, or third-party services, etc. AI management module 210 can select one or more selected global AI models 216 based on, for example, matching the definition of each global AI model (e.g., the associated tasks, sets of input-related attributes, and / or sets of output-related attributes defined for each global AI model) with the task request. AI management module 210 can select a single global AI model 216 or select multiple global AI models 216 to meet the task requirements (where each selected global AI model 216 can generate inference data for a subset of the task requirements).

[0105] After selecting a global AI model 216 for a task request, the AI ​​management module 210 performs training on the global AI model 216, for example, using global data from a global AI database 218 maintained by the AI ​​management module 210 (e.g., using training capabilities provided by AIMF 212). Training data from the global AI database 218 may include non-RT data (e.g., potentially exceeding milliseconds or one second) and may include network data and / or model data collected from one or more AI execution modules 220 managed by the AI ​​management module 210. After training is complete (e.g., the loss function for each global AI model 216 has converged), the selected global AI model 216 is executed to generate a set of global (or baseline) inference data (e.g., using model execution capabilities provided by AIMF 212). The global inference data may include global inference (or baseline) control parameters to be implemented at system node 120. The AI ​​management module 210 may also extract global model parameters (e.g., trained weights of the global AI model) from the trained global AI model for use by local AI models at the AI ​​execution module 220. The control parameters and / or global model parameters inferred globally are transmitted to the AI ​​execution module 220, for example, as configuration information in a configuration message (e.g., using the output function of AICF214).

[0106] At AI execution module 220, configuration information is received and optionally preprocessed (e.g., using the input functionality of AICF 224). The received configuration information may include model parameters, which AI execution module 220 uses to identify and configure one or more local AI models 226. For example, the model parameters may include an identifier of which local AI model 226 AI execution module 220 should select from a plurality of available local AI models 226 (e.g., the plurality of possible local AI models and their unique identifiers may be predefined by network standards or may be preconfigured at system node 120). The selected local AI model 226 may be similar to the selected global AI model 216 (e.g., having the same model definition and / or the same model identifier). The model parameters may also include globally trained weights, which may be used to initialize the weights of the selected local AI model 226. For example, based on a task request, a selected local AI model 226 (after being configured using model parameters received from the AI ​​management module 210) can be executed to generate one or more of the following inferred control parameters: mobility control, interference control, cross-carrier interference control, cross-cell resource allocation, RLC functions (e.g., ARQ, etc.), MAC functions (e.g., scheduling, power control, etc.) and / or PHY functions (e.g., RF and antenna operation, etc.).

[0107] The configuration information may also include control parameters (multiple) of inferred data generated based on the selected global AI model 216, which can be directly used to configure one or more control modules at system node 120. For example, the control parameters can be converted from the output format of the global AI model 216 (e.g., using the output functionality of AICF 224) into control instructions that the control modules recognize at system node 120. Control parameters from AI management module 210 can be adjusted or updated to generate locally inferred control parameters (e.g., using the model execution functionality provided by AIEF 222) by training a selected local AI model 226 on local network data. In an example where AI execution module 220 is implemented at system node 120, system node 120 may also transmit control parameters (whether received directly from AI management module 210 or generated using the selected local AI model 226) to one or more UEs 110 (not shown) served by system node 120.

[0108] System node 120 can also pass configuration information to one or more UEs 110 to configure UEs 110 to collect real-time or near-RT local network data. System node 120 can also configure itself to collect real-time or near-RT local network data. The local network data collected by UEs 110 and / or system node 120 can be stored in a local AI database 228 maintained by AI execution module 220 and used for near-RT training of a selected local AI model 226 (e.g., using the training function of AIEF 222). As previously described, training of the selected local AI model 226 can be performed relatively quickly (compared to training of the selected global AI model 216) to enable near-RT generation of inference data while collecting local data (to enable near-RT adaptation to dynamic real-world environments). For example, training of the selected local AI model 226 may involve fewer training iterations compared to training the selected global AI model 216. Alternatively, the trained parameters (e.g., trained weights) of the selected local AI model 226 after near-RT training on local network data can be extracted and stored as local model data in the local AI database 228.

[0109] In some examples, one or more control modules in the control modules at system node 120 (and optionally, one or more UEs 110 served by RAN 120) can be directly configured based on control parameters included in the configuration information from AI management module 210. In some examples, one or more control modules in the control modules at system node 120 (and optionally, one or more UEs 110 served by RAN 120) can be controlled based on locally inferred control parameters generated by the selected local AI model 226. In some examples, one or more control modules in the control modules at system node 120 (and optionally, one or more UEs 110 served by RAN 120) can be jointly controlled by control parameters from AI management module 210 and locally inferred control parameters.

[0110] Compared to a longer-term data storage at the global AI database 218, the local AI database 228 can be a shorter-term data storage (e.g., a cache or buffer). Local data maintained in the local AI database 228, including local network data and local model data, can be transferred (e.g., using the output functionality provided by AICF 224) to the AI ​​management module 210 for updating the global AI model 216.

[0111] At AI management module 210, local data collected from one or more AI execution modules 220 is received (e.g., using the input functionality provided by AICF 214) and added as global data to global AI database 218. This global data can be used for non-RT training of the selected global AI model 216. For example, if the local data from AI execution module 220 includes weights from the local training of a local AI model (if the local AI model has been updated via near-RT training), AI management module 210 can aggregate the locally trained weights and use the aggregation result to update the weights of the selected global AI model 216. After the selected global AI model 216 has been updated, it can be executed to generate updated global inference data. The updated global inference data (e.g., using the output functionality provided by AICF 214) can be transmitted to AI execution module 220, for example, as another configuration message or as an update message. In some examples, the update message transmitted to AI execution module 220 may only include control parameters or model parameters that have been changed from previous configuration messages. The AI ​​execution module 220 can receive and process updated configuration information in the manner described above.

[0112] exist FIG. 1J In the example shown, the AI ​​management module 210 performs continuous data collection, training of a selected global AI model 216, and execution of the trained global AI model 216 to generate updated data (including updated global inference control parameters and / or global model parameters) to continuously meet task requests (e.g., meeting one or more KPIs included as task requirements in the task request). The AI ​​execution module 220 can similarly perform continuous updates of configuration parameters, continuous collection of local network data, and optional continuous training of a selected local AI model 226 to continuously meet task requests (e.g., meeting one or more KPIs included as task requirements in the task request). FIG. 1J As shown, the collection of local network data, the training of global (or local) AI models, and the generation of updated inference data (whether global or local) can be performed repeatedly as a loop, for example, at least for the duration indicated in the task request (or until the task request is updated or replaced).

[0113] Now, let's describe another example of a collaborative task request. For example, a task request could be a request for the collaborative training of an AI model, which may include an identifier for the AI ​​model to be collaboratively trained, an identifier for and / or for collecting data used to train the AI ​​model, a dataset for training the AI ​​model, model parameters for collaboratively updating local training of the global AI model, and / or training objectives or requirements, etc. Task requests can be received from clients of wireless system 100E or 100F, from external network 150, and / or from nodes within wireless system 100E or 100F (e.g., from system node 120 itself).

[0114] At AI management module 210, upon receiving a task request, AI management module 210 executes functions (e.g., using functions provided by AIMF 212 and / or AICF 214) to perform initial setup and configuration according to the task request. For example, AI management module 210 can use the functions of AICF 214 to select and initialize one or more AI models according to the requirements of the collaborative task (e.g., according to the identifier of the AI ​​model to be collaboratively trained and / or according to the parameters of the AI ​​model to be collaboratively updated).

[0115] After selecting a global AI model 216 for a task request, the AI ​​management module 210 performs training on the global AI model 216. For co-training, the AI ​​management module 210 can use training data provided and / or identified in the task request to train the global AI model 216. For example, the AI ​​management module 210 can use model data (e.g., locally trained model parameters) collected from one or more AI execution modules 220 managed by the AI ​​management module 210 to update the parameters of the global AI model 216. In another example, the AI ​​management module 210 can use network data (e.g., locally generated and / or collected user data) collected from one or more AI execution modules 220 managed by the AI ​​management module 210 to train the global AI model 216 on behalf of the AI ​​execution modules 220. After training is complete (e.g., the loss function of each global AI model 216 has converged), model data extracted from the selected global AI model 216 (e.g., globally updated weights of the global AI model) can be transferred at the AI ​​execution module 220 for use by the local AI models. Global model parameters can be transmitted to AI execution module 220, for example, as configuration information in a configuration message (e.g., using the output function of AICF 214).

[0116] At AI execution module 220, configuration information includes model parameters that AI execution module 220 uses to update one or more corresponding local AI models 226 (e.g., the AI ​​model as the target of co-training, as identified in a co-training task request). For example, model parameters may include weights from global training, which can be used to update the weights of the selected local AI model 226. AI execution module 220 can then execute the updated local AI model 226. Additionally or alternatively, AI execution module 220 can continue to collect local data (e.g., local raw data and / or local model data), which can be maintained in local AI database 228. For example, AI execution module 220 can transfer newly collected local data to AI management module 210 to continue co-training.

[0117] At AI management module 210, local data collected from one or more AI execution modules 220 is received (e.g., using the input functionality provided by AICF 214), and this local data can be used to collaboratively update the selected global AI model 216. For example, if the local data from AI execution module 220 includes locally trained weights of a local AI model (if the local AI model has been updated via near-RT training), AI management module 210 can aggregate the locally trained weights and use the aggregation result to collaboratively update the weights of the selected global AI model 216. After the selected global AI model 216 has been updated, the updated model parameters can be transmitted back to AI execution module 220. This collaborative training, including communication between AI management module 210 and AI execution module 220, can continue until termination conditions are met (e.g., model parameters have sufficiently converged, the objective optimization and / or requirements of the collaborative training have been achieved, a timer has expired, etc.). In some examples, the requester of the collaborative task can send a message to AI management module 210 indicating that the collaborative task should end.

[0118] It can be noted that in some examples, the AI ​​management module 210 can participate in collaborative tasks without requiring detailed information about the data being used for training and / or the AI ​​model being collaboratively trained. For example, the requester of the collaborative task (e.g., system node 120 and / or UE 110) can define optimization objectives and / or identify the AI ​​model to be collaboratively trained, and can also identify and / or provide the data to be used for training. In some examples, a node acting as a public AI service center (or plug-in AI device) can, for example, implement the AI ​​management module 210 from a third party, which can provide the functionality of the AI ​​management module 210 (AI modeling and / or AI parameter training functions) based on relevant training data and / or task requirements from a request from a customer or system node 120 (e.g., BS) or UE 110. In this way, the AI ​​management module 210 can be implemented as a standalone and public AI node or device that can provide AI-specific functions for system node 120 or UE 110 (e.g., as an AI modeling training toolkit). However, the AI ​​management module 210 may not be directly involved in any wireless system control. If a wireless system desires or requires its specific control objectives to remain private or confidential, but requires the AI ​​modeling and training capabilities provided by the AI ​​management module 210 (for example, the AI ​​management module 210 does not even need to know about any AI execution module 220 present in the system node 120 or UE 110 that is requesting the task), then this implementation of the AI ​​management module 210 may be useful.

[0119] The following are some examples of how the AI ​​management module 210 collaborates with the AI ​​execution module 220 to fulfill task requests. It should be understood that these examples are not intended to be limiting. Furthermore, these examples are described in the context of implementing the AI ​​execution module 220 at system node 120. However, it should be understood that the AI ​​execution module 220 may be additionally or alternatively implemented at one or more UEs 110.

[0120] Example network task requests could be requests for low-latency services, such as URLLC traffic. AI management module 210 performs initial configuration based on this network task to set latency constraints (e.g., a maximum 2 ms delay in end-to-end communication). AI management module 210 also selects one or more global AI models 216 to handle this network task, such as selecting a global AI model associated with the URLLC. AI management module 210 trains the selected global AI model 216 using training data from global AI database 218. The trained global AI model 216 is executed to generate global inference data, including global control parameters that enable high-reliability communication (e.g., inference parameters for waveforms, inference parameters for interference control, etc.). AI management module 210 transmits configuration messages, including the globally inferred control parameters and model parameters, to AI execution module 220 at system node 120. AI execution module 220 outputs the received globally inferred control parameters to configure the appropriate control module at system node 120. AI execution module 220 also identifies and configures a local AI model 226 associated with the URLLC based on the model parameters. The local AI model 226 is executed to generate locally inferred control parameters for the control module at system node 120 (which can replace or be used in addition to globally inferred control parameters). For example, control parameters that may be inferred to satisfy the URLLC task may include parameters for: a fast handover switching scheme for URLLC, an interference control scheme for URLLC, defined cross-carrier resource allocation (to reduce cross-carrier interference), the RLC layer may not be configured with ARQ (to reduce latency), the MAC layer may be used to use unlicensed scheduling or conservative resource configuration with power control for uplink communication, and the PHY layer may be used for waveform and antenna configuration optimized for URLLC. The AI ​​execution module 220 collects local network data (e.g., channel status information (CSI), air link delay, end-to-end delay, etc.) and transmits the local data (which may include both collected local network data and local model data, such as the weights of the local training of the local AI model 226) to the AI ​​management module 210. AI management module 210 updates global AI database 218 and performs non-RT training on global AI model 216 to generate updated inference data. These operations can be repeated to continue fulfilling task requests (i.e., enabling URLLC).

[0121] Another example network task request could be a high-throughput request for file download. AI management module 210 performs an initial configuration based on this network task to set high-throughput requirements (e.g., high spectral efficiency for transmission). AI management module 210 also selects one or more global AI models 216 to handle this network task, such as global AI models related to spectral efficiency. AI management module 210 trains the selected global AI model 216 using training data from global AI database 218. The trained global AI model 216 is executed to generate global inference data, which includes global control parameters enabling high spectral efficiency (e.g., efficient resource scheduling, multi-TRP handover schemes, etc.). AI management module 210 transmits a configuration message, including the global inference control parameters and model parameters, to AI execution module 220 at system node 120. AI execution module 220 outputs the received global inference control parameters to configure an appropriate control module at system node 120. AI execution module 220 also identifies and configures local AI models 226 associated with spectral efficiency based on the model parameters. The local AI model 226 is executed to generate locally inferred control parameters for the control module at system node 120 (which can replace or be used in addition to the globally inferred control parameters). For example, control parameters that might be inferred to meet high throughput tasks could include parameters for: multi-TRP handover schemes, interference control schemes for model interference control, carrier aggregation and dual-connectivity multicarrier schemes, a fast ARQ configuration for the RLC layer, aggressive resource scheduling and power control for uplink communication for the MAC layer, and antenna configuration for massive MIMO for the PHY layer. The AI ​​execution module 220 collects local network data (e.g., actual throughput) and transmits the local data (which may include both collected local network data and local model data, such as the weights of the local training of the local AI model 226) to the AI ​​management module 210. The AI ​​management module 210 updates the global AI database 218 and performs non-RT training on the global AI model 216 to generate updated inference data. These operations can be repeated to continue meeting task requests (i.e., enabling high throughput).

[0122] FIG. 2 A signaling diagram illustrating an exemplary communication process 200 according to some exemplary embodiments of the present disclosure is shown. Reference will be made to this diagram for discussion purposes only. FIG. 1A to 1J The communication process 200 is described. The communication process 200 may involve a first network device (e.g., such as...). FIG. 1B and 2 RAN 120a shown), and the second network device (e.g., such as...) FIG. 1B and 2RAN120b shown), and terminal equipment (e.g., such as FIG. 1B and 2 (UE 110a shown).

[0123] like FIG. 2 As shown, the first network device 120a sends (210) a first request 201 to the second network device 120b, which instructs the second network device to provide a first AI / ML model. On the other side of the communication, the second network device 120b receives (212) the first request 201 from the first network device 120a. Then, at the second terminal device 120b, in box 215, the second terminal device 120b, based on the pre-trained AI / ML model at the second network device 120b (e.g., such as...),... FIG. 1I The pre-trained AI / ML model shown in the diagram generates an AI / ML model. Here, the pre-trained AI / ML model is pre-trained for multiple tasks, including tasks that the first network device 120a will perform using the generated AI / ML model. Then, the second network device 120b sends (220) the first AI / ML model 202 (i.e., the AI / ML model generated in box 215) to the first network device 120a. On the other side of the communication, the first network device 120a receives (222) the first AI / ML model 202 from the second network device 120b. At box 225, the first network device 120a obtains a fine-tuned AI / ML model based on the first AI / ML model 202. In this way, a relatively lightweight custom local AI / ML model can be obtained from a rather large (and “bulky”) global base model at the CN node or a third party, thereby reducing the training complexity at the first network device.

[0124] For example, to obtain a fine-tuned AI / ML model, the first network device 120a can collect local data and then fine-tune the first AI / ML model 202 based on that data. In this way, a more accurate local AI / ML model can be obtained to perform the task at the first network device 120a. Alternatively or additionally, the first network device 120a can, for example, send data to the second network device 120b before receiving the first AI / ML model from the second network device 120b. In this case, the second network device 120b can fine-tune the pre-trained AI / ML model based on the data to obtain a fine-tuned AI / ML model, and then send the fine-tuned AI / ML model to the first network device 120a as such. FIG. 2 The first AI / ML model 202 is shown in the figure. In this way, the AI / ML model 202 obtained from the second network device 120b is more accurate for the first network device 120a, and training time at the first network device 120a can be saved.

[0125] The first network device 120a can also send a second request to the second network device 120b, instructing the second network device 120b to provide a second AI / ML model. Upon receiving the second request, the second network device 120b can generate a second AI / ML model based on a pre-trained AI / ML model and send the second AI / ML model to the first network device 120a. On the other side of the communication, the first network device 120a receives the second AI / ML model from the second network device 120b. In this way, the first network device 120a can obtain more than one AI / ML model to perform various tasks. For example, at the first network device 120a, the second request can be sent together with the first request, and the second AI / ML model can be received together with the first AI / ML model 202. At the second network device 120b, the second request can be received together with the first request, and the second AI / ML model can be sent together with the first AI / ML model. This reduces signaling overhead compared to sending the two requests separately. In this scenario, at the second network device 120b, to send the second AI / ML model along with the first AI / ML model to the first network device 120a, the second terminal device 120b can send at least one model parameter common to both the first and second AI / ML models, at least one model parameter specific to the first AI / ML model 202, and at least one model parameter specific to the second AI / ML model. On the other side of the communication, at the first network device 120a, to receive the second AI / ML model along with the first AI / ML model 202, the first network device 120a can receive at least one model parameter common to both the first and second AI / ML models, at least one model parameter specific to the first AI / ML model, and at least one model parameter specific to the second AI / ML model. By sending / receiving the common at least one model parameter once instead of twice—that is, separately for the first AI / ML model 202 and the second AI / ML model—signaling overhead can be reduced.

[0126] The first network device 120a, one of the other (or more) first network devices, can also send a fine-tuned AI / ML model to the second network device 120b. On the other side of the communication, the second network device 120b can receive at least one AI / ML model, including the AI / ML model provided by the first network device 120a, from at least one network device including the first network device 120a. The second network device 120b can then generate an updated AI / ML model based on the at least one AI / ML model and send the updated AI / ML model to the first network device 120a. On the other side of the communication, the first network device 120a can receive the updated AI / ML model from the second network device 120b. In this way, the AI / ML model used at the first network device 120a can perform tasks more accurately.

[0127] Now return to the reference FIG. 2 The first network device 120a can also send (230) fine-tuned AI / ML models (e.g., such as...) to at least one terminal device including terminal device 110a. FIG. 2 The AI / ML model 203 shown is illustrated. On the other side of the communication, for example, terminal device 110a can receive (232) AI / ML model 203 from first network device 120a. Then, at block 235, terminal device 110a can fine-tune AI / ML model 203 based on data collected at terminal device 110a to obtain an updated fine-tuned AI / ML model, and then send (240) the updated AI / ML model to first network device 120a (e.g., as shown). FIG. 2 The AI / ML model 204 shown is illustrated. On the other side of the communication, the first network device 120a can receive (242) the updated AI / ML model 204 from the terminal device 110a and generate an updated AI / ML model based on the received updated AI / ML model 204. In this way, the first network device 120a can fine-tune the AI / ML model it uses to perform tasks more accurately, especially tasks related to the terminal device (e.g., terminal device 110a). It should be noted that the fine-tuning performed by the terminal device (e.g., terminal device 110a) is an optional solution to the solution proposed in this disclosure, and therefore in FIG. 2 The middle part is represented by a dashed line.

[0128] In a more specific example, at least one terminal device may include multiple terminal devices (including terminal device 110a), and at least one third AI / ML model may include multiple AI / ML models. In this case, in order to generate an updated AI / ML model based on at least one third AI / ML model, the first network device 120a may aggregate multiple AI / ML models to generate a fourth AI / ML model as the updated AI / ML model. In this way, the AI / ML model at the first network device 120a can perform a specific task more accurately, especially when the task is related to multiple terminal devices (including terminal device 110a).

[0129] More specifically, in order to generate an updated AI / ML model based on at least one AI / ML model, the first network device 120a can aggregate a fourth AI / ML model and a fine-tuned AI / ML model to generate a fifth AI / ML model as the updated AI / ML model. In this way, the AI / ML model at the first network device 120a can perform specific tasks more accurately, especially when the task is related to at least one terminal device (including terminal device 110a).

[0130] Then, the first network device 120a, one of the other (or more) first network devices, can send an updated AI / ML model to the second network device 120b. On the other side of the communication, the second network device 120b can receive at least one AI / ML model from at least one network device including the first network device 120a, the at least one AI / ML model including the updated AI / ML model provided by the first network device 120a. Here, as described above, an updated AI / ML model is generated based on at least one AI / ML model provided by at least one terminal device. Then, the second terminal device 120b can generate a further updated AI / ML model based on the at least one AI / ML model and send the further updated AI / ML model to the first network device 120a. On the other side of the communication, the first network device 120a can receive the further updated AI / ML model from the second network device 120b. In this way, the AI / ML model at the first network device 120a can more accurately perform a specific task, especially when the task is related to at least one terminal device including terminal device 110a.

[0131] By fine-tuning and updating the AI / ML model in various ways as described above, the first network device 120a can perform tasks using at least one of the following: an AI / ML model, a fine-tuned AI / ML model, an updated AI / ML model, or a further updated AI / ML model. Alternatively, the first network device 120a can send data related to at least one of a plurality of tasks and stored in an AI / ML database at the first network device 120a to the second network device 120b. In this case, the second network device 120b can receive data related to at least one of the plurality of tasks and stored at the first network device 120a from the first network device 120a, and store the received data at the second network device 120b. In this way, the global base model at the second network device 120b can be trained with the received data to generate a custom AI / ML model specific to the first network device before being sent to the first network device 120a. More specifically, the AI / ML model sent to the first network device 120a can be a fine-tuned AI / ML model based on data from the first network device 120a. In other words, before sending the AI / ML model to the first network device 120a, the second network device 120b can perform fine-tuning to obtain a custom AI / ML model specifically for the first network device 120a. In this way, the first network device 120a can perform local tasks using a more accurate AI / ML model.

[0132] In this way, according to communication process 200, a relatively lightweight custom local AI / ML model can be obtained at the first network device 120a from the rather large (and "bulky") global base model at the second network device 120b, thereby reducing the training complexity at the first network device 120a. Simultaneously, the AI / ML model obtained at the first network device 120a can perform tasks more accurately at the first network device 120a.

[0133] FIG. 3 A schematic diagram of an exemplary AI model implementation 300 according to some embodiments of the present disclosure is shown. FIG. 3As shown, the AI ​​model can be implemented in various locations. For example, the AI ​​model can be implemented at over-the-top (OTT), Edge, BS, or UE. The AI ​​model at OTT is in the application layer of the Open System Interconnection (OSI) model, the AI ​​model at Edge is in the Packet Data Unit (PDU) layer, and the AI ​​model at RAN (BS or UE) can be in the Service Data Adaptation Protocol (SDAP) layer, Packet Data Convergence Protocol (PDCP) layer, Radio Link Control (RLC) layer, Media Access Control (MAC) layer, or Physical (PHY) layer.

[0134] FIG. 4 A signaling diagram illustrating another exemplary communication process 400 according to some embodiments of this disclosure is shown. In communication process 400, the RAN node receives at least one custom AI model from the core network (CN) or a third party, and performs fine-tuning on the custom AI model at the RAN node. In this case, the AI ​​Execution Function (AIEF) is located in the RAN node, and the AI ​​Management Function (AIMF) is located in the CN or the third party. Reference will be made to this diagram for discussion purposes only. FIG. 1A to 1J and FIG. 2 Describe the communication process 400.

[0135] exist FIG. 4 In the context of the global AI model (basic model) 403, it is as follows: FIG. 1I The example shown is a pre-trained large model 100I, implemented in a core network or third-party 401, which is as follows: FIG. 2 The example shown is the second network device 120b. A global AI database (DB) 402 for the global AI model 403 is also implemented in the core network or a third party 401. On the RAN side, there is a network device 404. Network device 404 can be, for example, a transmit and receive point (TRP). Local AI models are deployed at a base station (BS) 405 (the base station is, for example,...). FIG. 2(Example of the first network device 120a shown). Base station 505 is directly or indirectly connected to network device 404. A local AI database 406 also exists for a local AI model 407 implemented at base station 405. For clarity, local AI model 407 can be scaled up to model 408. (And...) FIG. 1I In comparison, it is clear that model 408 at base station 405 is superior to the pre-trained large model (here, in...). FIG. 4 The global AI model (403) is smaller and simpler. This is because at base station 405, the task to be performed (here, in...) FIG. 4 The number of tasks (task-1 and task-2) is far less than the number of tasks used to pre-train the global AI model.

[0136] Specifically, at 410, base station 405, acting as a RAN node, sends one or more task requests to CN or third party 401, requesting CN or third party 401 to provide a corresponding AI model to base station 405. Upon receiving one or more task requests, CN or third party 401 generates one or more custom AI models and sends these custom AI models to base station 405 at 415. Here, if CN or third party 401 sends more than one custom AI model to base station 405, some parameters of the custom AI models may differ slightly. FIG. 4 As shown, the model parameters for Task-1 and Task-2 are different within a single layer. Therefore, within these different layers, the indicator could be that the model parameter for Task-1 is weight-1, and the model parameter for Task-2 is weight-2. For other parameters, one indicator is sufficient because the parameters for both tasks are the same and can be used interchangeably.

[0137] At 420, base station 405 collects training data, for example, through transmission reception point (TRP) sensing or TRP measurement. Base station 405 can store the collected training data in a local AI database 406. Then, at 425, base station 405 can use the collected training data to fine-tune one or more custom AI models received from the CN or a third party 401. Through fine-tuning, a fine-tuned AI model can be obtained from the custom AI model, and base station 405 can use the fine-tuned AI model to perform local AI tasks, such as... FIG. 4 Task-1 and / or Task-2 are shown.

[0138] Optionally, at 490, base station 405 can send locally collected training data from local AI database 406 to global AI database 402, so that CN or third party 401 can use, for example, the training data to fine-tune and update global AI model 403.

[0139] In this way, a relatively lightweight custom local AI / ML model 407 can be obtained from the rather large (and “bulky”) global base model 403 at CN or a third party 401, thereby reducing the training complexity at base station 405. At the same time, the local AI / ML model 407 at base station 405 is more accurate, so base station 405 can perform tasks more accurately.

[0140] FIG. 5 A signaling diagram illustrating another exemplary communication process 500 according to some embodiments of this disclosure is shown. In communication process 500, a random access network (RAN) node receives at least one custom AI model from a core network (CN) or a third party, and performs fine-tuning on the custom AI model at a RAN node including BS and UE nodes. In this case, the AI ​​Execution Function (AIEF) is located in the RAN node, and the AI ​​Management Function (AIMF) is located in the CN or a third party. Reference will be made to this diagram for discussion purposes only. FIG. 1A to 1J , FIG. 2 and FIG. 4 Describe the communication process 500.

[0141] exist FIG. 5 In the context of the global AI model (basic model) 503, it is as follows: FIG. 1I The example shown is a pre-trained large model 100I, implemented in the core network or a third party 501. A global AI database (DB) 502, used for the global AI model 503, is also implemented in the core network or a third party 501. On the RAN side, there is a network device 504. (The last sentence appears to be incomplete and possibly contains errors.) FIG. 4 Similar to network device 404 shown, network device 504 can be a transmit and receive point (TRP). The local AI model is deployed at a base station (BS) 505. Base station 505 is directly or indirectly connected to network device 404. A local AI database 506 also exists for the local AI model 507 implemented at base station 505. Other databases also exist, for example... FIG. 1B UE110a and UE110b are shown. FIG. 5 The following components can be integrated: central and global AI model 503, core network or third-party 501, global AI database 502, network equipment 504, base station 505, local AI database 506, and local AI model 507. FIG. 4The global AI model 403, core network or third party 401, global AI database 402, network device 404, base station 405, local AI database 406, and local AI model 407 shown are similar or identical. The communication process 500 differs from... FIG. 4 The main difference between the communication process 400 shown and the previous one is that in communication process 500, the UE node also participates in fine-tuning a custom AI model obtained from the CN or a third party 501. FIG. 5 The operations at points 510, 515, 520, 525, and 590 are similar to... FIG. 4 The operations at positions 410, 415, 420, 425, and 490 are very similar.

[0142] Specifically, at 510, base station 505, acting as a RAN node, sends one or more task requests to CN or third party 501, requesting CN or third party 501 to provide a corresponding AI model to base station 505. Upon receiving one or more task requests, CN or third party 501 generates one or more custom AI models and sends these custom AI models to base station 505 at 515. Here, if CN or third party 501 sends more than one custom AI model to base station 505, some parameters of the custom AI models may differ slightly. FIG. 4 As shown, the model parameters for Task-1 and Task-2 are different within a single layer. Therefore, within these different layers, the indicator could be that the model parameter for Task-1 is weight-1, and the model parameter for Task-2 is weight-2. For other parameters, one indicator is sufficient because the parameters for both tasks are the same and can be used interchangeably.

[0143] At 520, base station 505 collects training data, for example, through transmission reception point (TRP) sensing or TRP measurement. Base station 505 can store the collected training data in a local AI database 506. Then, at 525, base station 505 can use the collected training data to fine-tune one or more custom AI models received from CN or a third party 501 to obtain a fine-tuned AI model.

[0144] Then, at 530, base station 505 further sends the local model from the BS side to one or more UEs to further fine-tune the local model, enabling multiple UEs (here, UE 110a and UE 110b) to train on the same or different tasks indicated by the base station at 530. At 535, each of the multiple UEs trains the AI ​​model received from base station 505 using its own data and reports the updated local AI model to base station 505. For example... FIG. 5As shown, UE 110a is the AI ​​model received during training for Task-1, and UE 100b is the AI ​​model received during training for Task-1 and / or Task-2. In this way, for example, gradient information can be updated to ensure data privacy.

[0145] At base station 505, at 540, base station 505 aggregates local AI models reported from multiple UEs to obtain an updated local AI model at base station 505. In one example, base station 505 may aggregate local AI models reported from multiple UEs to generate an updated local AI model. In another example, base station 505 may aggregate local AI models reported from multiple UEs and a fine-tuned AI model at base station 505 to generate an updated local AI model.

[0146] The process from 520 to 540 can be repeated several times until a (pre)configured or (pre)defined condition is met. As an example, the condition could be that the difference between a first output corresponding to one or more inputs of the updated local AI model and a second input corresponding to one or more of the same inputs of the global AI model 503 is less than a (pre)configured or (pre)defined threshold.

[0147] Through fine-tuning, a finely tuned AI model can be obtained from a custom AI model. Base station 505 can then use this finely tuned AI model to perform localized AI tasks, such as... FIG. 4 Task-1 and / or Task-2 are shown.

[0148] Optionally, at 590, base station 505 can send locally collected training data from local AI database 506 to global AI database 502, so that CN or a third party 501 can use, for example, the training data to fine-tune and update global AI model 503.

[0149] In this way, a relatively lightweight custom local AI / ML model 507 can be obtained from the rather large (and “bulky”) global base model 503 at CN or a third party 501, thereby reducing the training complexity at base station 505. At the same time, the local AI / ML model 507 at base station 505 is more accurate, so base station 505 can perform tasks more accurately (especially tasks related to UE nodes involved in fine-tuning the local AI model 507).

[0150] FIG. 6A signaling diagram illustrating another exemplary communication process 600 according to some embodiments of this disclosure is shown. In communication process 600, a random access network (RAN) node receives at least one custom AI model from a core network (CN) or a third party. Fine-tuning of the custom AI model is performed at the RAN node and the CN (or the third party). In this case, the AI ​​Execution Function (AIEF) is located in the RAN node, and the AI ​​Management Function (AIMF) is located in the CN or the third party. Reference will be made to this diagram for discussion purposes only. FIG. 1A to 1J , FIG. 2 and FIG. 4 Describe the communication process 600.

[0151] exist FIG. 6 In the context of the global AI model (basic model) 603, it is as follows: FIG. 1I The example shown is a pre-trained large model 100I, implemented in the core network or a third party 601. A global AI database (DB) 602 for the global AI model 603 is also implemented in the core network or a third party 601. On the RAN side, there is a network device 604. (The last sentence appears to be incomplete and possibly contains errors.) FIG. 4 Similar to network device 404 shown, network device 604 can be a transmit and receive point (TRP). Local AI model 607 is implemented at base station (BS) 605. Base station (BS) 605 is directly or indirectly connected to network device 604. A local AI database 606 also exists for the local AI model 407 implemented at base station 605. FIG. 6 In this context, the global AI model 603, the core network or third party 601, the global AI database 602, network equipment 604, base stations 605, the local AI database 606, and the local AI model 607 can be integrated with, for example... FIG. 4 The global AI model 403, core network or third party 401, global AI database 402, network device 404, base station 405, local AI database 406, and local AI model 407 shown are similar or identical. The communication process 600 differs from... FIG. 4 The main difference in the communication process 400 shown is that, in communication process 600, base station 605 reports its local AI model 607 to CN or a third party 601. The reported AI models are aggregated to obtain an updated local AI model for base station 605, which is then sent to base station 605 for future use. FIG. 5The operations at points 610, 615, 620, 625, and 690 are similar to... FIG. 4 The operations at positions 410, 415, 420, 425, and 490 are very similar.

[0152] Specifically, at 610, base station 605, acting as a RAN node, sends one or more task requests to CN or third party 601, requesting CN or third party 601 to provide a corresponding AI model to base station 605. Upon receiving one or more task requests, CN or third party 601 generates one or more custom AI models and sends these custom AI models to base station 605 at 615. Here, if CN or third party 601 sends more than one custom AI model to base station 605, some parameters of the custom AI models may differ slightly. FIG. 4 As shown, the model parameters for Task-1 and Task-2 are different within a single layer. Therefore, within these different layers, the indicator could be that the model parameter for Task-1 is weight-1, and the model parameter for Task-2 is weight-2. For other parameters, one indicator is sufficient because the parameters for both tasks are the same and can be used interchangeably.

[0153] At 620, base station 605 collects training data, for example, through TRP sensing or TRP measurement. Base station 605 can store the collected training data in a local AI database 606. Then, at 625, base station 605 can use the collected training data to fine-tune one or more custom AI models received from the CN or a third party 601. Through fine-tuning, a fine-tuned AI model can be obtained from the custom AI models, and base station 605 can use the fine-tuned AI model to perform local AI tasks, such as... FIG. 4 Task-1 and / or Task-2 are shown.

[0154] Optionally, at 690, base station 605 can send locally collected training data from local AI database 606 to global AI database 602, so that CN or a third party 601 can use, for example, the training data to fine-tune and update global AI model 603.

[0155] At 630, base station 605 sends the fine-tuned local AI model 607 to CN or third party 601. During the fine-tuning process, multiple BS nodes (including base station 605) can participate in training, and multiple BS nodes send their fine-tuned local models (including local AI model 607) to CN or third party 601. Therefore, at 635, CN or third party 601 aggregates the local AI models reported from one or more BS nodes and obtains an updated task-specific AI model. For example, CN or third party 601 can directly aggregate the reported local AI models to generate an updated AI model. Alternatively, CN or third party 601 can first generate a custom AI model specifically for the task requested by the task request received from one or more BS nodes, then at 690 use the reported data to fine-tune the custom AI model to generate a locally fine-tuned AI model, and finally aggregate the reported local AI model and the locally fine-tuned AI model to generate an updated AI model. As an alternative, CN or third party 601 can first use the reported training data at 690 to fine-tune and update the global AI model 603, then generate a custom AI model from the updated global AI model specifically for the task requested by a task request received from one or more BS nodes, and finally aggregate the reported local AI model and the custom AI model to generate an updated AI model. At 640, CN or third party 601 sends the updated AI model to base station 605.

[0156] The process from 620 to 640 can be repeated several times until a (pre)configured or (pre)defined condition is met. As an example, the condition may be that the difference between a first output corresponding to one or more inputs of the updated local AI model 607 and a second input corresponding to one or more of the same inputs of the global AI model 603 is less than a (pre)configured or (pre)defined threshold.

[0157] Through fine-tuning performed by the BS node (here, base station 605) and the CN or a third party 601, an updated, fine-tuned AI model can be obtained, and base station 605 can use the AI ​​model to perform local AI tasks, such as... FIG. 4 Task-1 and / or Task-2 are shown.

[0158] In this way, a relatively lightweight custom local AI / ML model 607 can be obtained from the rather large (and “bulky”) global base model 603 at CN or a third party 601, thereby reducing the training complexity at base station 605. At the same time, the local AI / ML model 607 at base station 605 is more accurate, so base station 605 can perform the task more accurately.

[0159] FIG. 7A signaling diagram illustrating another exemplary communication process 700 according to some embodiments of this disclosure is shown. In communication process 700, the RAN node receives at least one custom AI model from the core network (CN) or a third party, and performs fine-tuning on the custom AI model at the RAN node, which includes BS and UE nodes. In this case, the AI ​​Execution Function (AIEF) is located in the RAN node, and the AI ​​Management Function (AIMF) is located in the CN or a third party. Reference will be made to this diagram for discussion purposes only. FIG. 1A to 1J , FIG. 2 and FIG. 4 Describe the communication process 700.

[0160] exist FIG. 7 In the context of the global AI model (basic model) 703, it is as follows: FIG. 1I The example shown is a pre-trained large model 100I, implemented in the core network or a third-party 701. A global AI database (DB) 702, used for the global AI model 703, is also implemented in the core network or a third-party 701. On the RAN side, there is a network device 704. (The last sentence appears to be incomplete and possibly contains errors.) FIG. 4 Similar to network device 404 shown, network device 704 can be a transmit and receive point (TRP). Local AI model 707 is deployed at base station (BS) 705. Base station 705 is directly or indirectly connected to network device 704. A local AI database 706 also exists for the local AI model 707 implemented at base station 705. Other databases also exist, for example... FIG. 1B UE 110a and UE 110b are shown. FIG. 7 Can be regarded as FIG. 5 and FIG. 6 The combination. In FIG. 7 In this context, the global AI model 503, the core network or third party 501, the global AI database 502, network devices 504, base stations 505, the local AI database 506, the local AI model 507, and UE 110a and UE 110b can interact with, for example... FIG. 5 The global AI model 503, core network or third party 501, global AI database 502, network device 504, base station 505, local AI database 506, and local AI model 507 shown are similar to or the same as UE 110a and UE 110b. The communication process 700 differs from... FIG. 5The main difference in the communication process 500 shown is that, in communication process 700, base station 705 reports its local AI model 607 to CN or a third party 701. The reported AI models are aggregated to obtain an updated local AI model for base station 705, which is then sent to base station 705 for future use. FIG. 5 The operations at points 710, 715, 720, 725, 730, 735, 740, and 790 are related to... FIG. 5 The operations at 510, 515, 520, 525, 530, 535, 540, and 490 are very similar.

[0161] Specifically, at 710, base station 705, acting as a RAN node, sends one or more task requests to CN or third party 701, requesting CN or third party 701 to provide a corresponding AI model to base station 705. Upon receiving one or more task requests, CN or third party 701 generates one or more custom AI models and sends these custom AI models to base station 705 at 715. Here, if CN or third party 701 sends more than one custom AI model to base station 705, some parameters of the custom AI models may differ slightly. FIG. 4 As shown, the model parameters for Task-1 and Task-2 are different within a single layer. Therefore, within these different layers, the indicator could be that the model parameter for Task-1 is weight-1, and the model parameter for Task-2 is weight-2. For other parameters, one indicator is sufficient because the parameters for both tasks are the same and can be used interchangeably.

[0162] At 720, base station 705 collects training data, for example, through TRP sensing or TRP measurement. Base station 705 can store the collected training data in a local AI database 706. Then, at 725, base station 705 can use the collected training data to fine-tune one or more custom AI models received from CN or a third party 701 to obtain a fine-tuned AI model.

[0163] Then, at 730, base station 705 further sends the local model from the BS side to one or more UEs to further fine-tune the local model, allowing multiple UEs (here, UE 110a and UE 110b) to train on the same or different tasks indicated by the base station at 730. At 735, each of the multiple UEs trains the AI ​​model received from base station 705 using its own data and reports the updated local AI model to base station 705. FIG. 7As shown, UE 110a is the AI ​​model received during training for Task-1, and UE 100b is the AI ​​model received during training for Task-1 and / or Task-2. In this way, for example, gradient information can be updated to ensure data privacy.

[0164] Optionally, at 790, base station 705 can send locally collected training data from local AI database 706 to global AI database 702, so that CN or a third party 701 can use, for example, the training data to fine-tune and update global AI model 703.

[0165] At base station 705, at 740, base station 705 aggregates local AI models reported from multiple UEs to obtain an updated local AI model at base station 705. In one example, base station 705 may aggregate local AI models reported from multiple UEs to generate an updated local AI model. In another example, base station 705 may aggregate local AI models reported from multiple UEs and a fine-tuned AI model at base station 705 to generate an updated local AI model.

[0166] Then, at 750, base station 705 sends the updated local AI model 607 (updated with the assistance of multiple UEs, including UE 110a and UE 110b) to CN or third party 701. During fine-tuning, multiple BS nodes (including base station 705) can participate in training, with each BS node sending its fine-tuned local model (including local AI model 707) to CN or third party 701. Therefore, at 755, CN or third party 701 aggregates the local AI models reported from one or more BS nodes and obtains an updated task-specific AI model. For example, CN or third party 701 can directly aggregate the reported local AI models to generate an updated AI model. Alternatively, CN or third party 701 can first generate a custom AI model specifically for the task requested by a task request received from one or more BS nodes, then fine-tune the custom AI model at 790 using the reported data to generate a locally fine-tuned AI model, and finally aggregate the reported local AI model and the locally fine-tuned AI model to generate an updated AI model. As an alternative, the CN or third party 701 can first use the reported training data at 790 to fine-tune the global AI model 703 to generate a custom AI model specifically for the task requested by a task request received from one or more BS nodes. Then, the reported local AI model and the custom AI model are aggregated to generate an updated AI model. At 760, the CN or third party 701 sends the updated AI model to the base station 705.

[0167] The process from 720 to 760 can be repeated several times until a (pre)configured or (pre)defined condition is met. As an example, the condition could be that the difference between a first output corresponding to one or more inputs of the updated local AI model and a second input corresponding to one or more of the same inputs of the global AI model 703 is less than a (pre)configured or (pre)defined threshold.

[0168] Through fine-tuning, a finely tuned AI model can be obtained from a custom AI model. The base station 705 can then use this finely tuned AI model to perform local AI tasks, such as... FIG. 4 Task-1 and / or Task-2 are shown.

[0169] In this way, a relatively lightweight custom local AI / ML model 707 can be obtained from the rather large (and “bulky”) global base model 703 at CN or a third party 701, thereby reducing the training complexity at base station 705. At the same time, the local AI / ML model 707 at base station 705 is more accurate, so base station 705 can perform tasks more accurately (especially tasks related to UE nodes involved in fine-tuning the local AI model 707).

[0170] FIG. 8 A signaling diagram illustrating another exemplary communication process 800 according to some exemplary embodiments of this disclosure is shown. In communication process 800, the RAN node sends a task request and its partial data to a CN or a third party, and fine-tuning of the AI ​​model is performed at the CN or third party using data reported from the RAN node. After fine-tuning, the CN or third party sends a custom model to the RAN node. In this case, the AI ​​Execution Function (AIEF) resides in the RAN node, and the AI ​​Management Function (AIMF) resides in the CN or third party. Reference will be made to this diagram for discussion purposes only. FIG. 1A to 1J , FIG. 2 and FIG. 4 Describe the communication process 800.

[0171] exist FIG. 8 In the context of the global AI model (basic model) 803, it is as follows: FIG. 1I The example shown is a pre-trained large model 100I, implemented in the core network or a third-party 801. A global AI database (DB) 802, used for the global AI model 803, is also implemented in the core network or a third-party 801. On the RAN side, there is a network device 804. (The last sentence appears to be incomplete and possibly refers to a different implementation.) FIG. 4Similar to network device 404 shown, network device 704 can be a transmit and receive point (TRP). Local AI model 807 is deployed at base station (BS) 805. Base station 805 is directly or indirectly connected to network device 704. A local AI database 806 also exists for the local AI model 807 implemented at base station 805. For clarity, local AI model 807 can be scaled up to model 808. (Similar to...) FIG. 1I In comparison, it is clear that model 808 at base station 805 is superior to the pre-trained large model (here, in...). FIG. 8 The global AI model (803) is smaller and simpler. This is because at base station 805, the task to be performed (here, in...) FIG. 8 The number of tasks (task-1 and task-2) is far less than the number of tasks used to pre-train the global AI model 803.

[0172] Specifically, at 810, base station 805 collects training data, for example, through TRP sensing or TRP measurement. Base station 805 can store the collected training data in a local AI database 806. UE 110a connected to base station 805 also provides (transmits) local data at UE 110a to base station 805 for training a local AI model to be used at base station 805. Base station 805 can also store reported data collected by UE 110a in the local AI database 806. Then, at 815, base station 805, as a RAN node, sends one or more task requests to CN or third party 801 to request CN or third party 801 to provide the corresponding AI model to base station 805. The processing order of 810 (or 812) and 815 can be interchanged.

[0173] At 820, base station 805 sends (reports) local training data from local AI database 806 to global AI database 802, allowing CN or third party 801 to use the training data to, for example, fine-tune global AI model 803 to generate a custom AI model for base station 805, i.e., a sub-model of global AI model 803, for future use by base station 803. At 825, CN or third party 801 performs fine-tuning on global AI model 803 using the training data reported from base station 805 (now stored in global AI database 802). In doing so, at 830, CN or third party 801 generates a custom AI model and sends the custom AI model to base station 805. Here, if CN or third party 801 sends more than one custom AI model to base station 805, some parameters of the custom AI models may have minor differences. FIG. 4As shown, the model parameters for Task-1 and Task-2 are different within a single layer. Therefore, within these different layers, the indicator could be that the model parameter for Task-1 is weight-1, and the model parameter for Task-2 is weight-2. For other parameters, one indicator is sufficient because the parameters for both tasks are the same and can be used interchangeably.

[0174] In this way, a relatively lightweight custom local AI / ML model 807 can be obtained from the rather large (and "bulky") global base model 803 at CN or a third party 801, thereby reducing the training complexity at base station 805. At the same time, the local AI / ML model 807 at base station 805 is more accurate, so base station 805 can perform tasks more accurately.

[0175] FIG. 9 A flowchart illustrating an exemplary method 900 implemented at a first network device according to some other embodiments of the present disclosure is shown. Reference will be made to this flowchart for discussion purposes. FIG. 1B , FIG. 2 and FIG. 4 to 8 Method 1000 is described from the perspective of the first network device 120a.

[0176] In box 910, the first network device 120a communicates with the second network device (e.g., ...). FIG. 2 The second network device 120b shown, or FIG. 4 The CN or third party shown in the image (401) sends the first request (e.g., as shown in the image). FIG. 2 The first request 201 shown, or as... FIG. 4 As shown in the task request at 410, the first request instructs the second network device to provide a first AI / ML model. At box 920, the first network device 120a receives the first AI / ML model from the second network device (e.g., as shown in the task request at 410). FIG. 2 The first AI / ML model 202 is shown in the diagram. At box 930, the first network device 120a obtains a fine-tuned AI / ML model based on the first AI / ML model.

[0177] In some example embodiments, to obtain a fine-tuned AI / ML model, the first network device 120a may perform fine-tuning on the first AI / ML model based on data from the first network device 120a. For example, the first network device 120a may collect data through TRP sensing or TRP measurement and store the collected data in a local AI database (e.g., such as...). FIG. 4 The local AI database 406 shown is used to fine-tune the first AI / ML model using the collected data to obtain a fine-tuned AI / ML model. In this way, a more accurate local AI / ML model can be obtained.

[0178] In some example embodiments, the first network device 120a may also send a second request to the second network device instructing the second network device to provide a second AI / ML model, and receive the second AI / ML model from the second network device. In this way, more than one task-specific AI / ML model can be obtained from the second network device.

[0179] In some example implementations, the second request can be sent together with the first request, and the second AI / ML model can be received together with the first AI / ML model. This reduces signaling overhead compared to sending the two requests separately.

[0180] In some example embodiments, when receiving the second AI / ML model together with the first AI / ML model, the first network device 120a receives at least one model parameter common to both the first and second AI / ML models, at least one model parameter specific to the first AI / ML model, and at least one model parameter specific to the second AI / ML model. By receiving the common at least one model parameter once instead of twice, i.e., separately for the first and second AI / ML models, signaling overhead can be reduced.

[0181] In some exemplary embodiments, the first network device 120a may also send a fine-tuned AI / ML model to the second network device and receive an updated AI / ML model from the second network device. In this way, the AI / ML model at the first network device 120a can perform specific tasks more accurately.

[0182] In some exemplary embodiments, the first network device 120a may also send a fine-tuned AI / ML model to at least one terminal device and receive at least one third AI / ML model from at least one terminal device. The first network device 120a may then generate an updated AI / ML model based on the at least one third AI / ML model. In this way, the AI / ML model at the first network device 120a can perform specific tasks more accurately, especially when the task is related to at least one terminal device.

[0183] In some exemplary embodiments, at least one terminal device includes multiple terminal devices, and at least one third AI / ML model includes multiple AI / ML models. In this case, in order to generate an updated AI / ML model based on at least one third AI / ML model, the first network device 120a may aggregate multiple AI / ML models to generate a fourth AI / ML model as the updated AI / ML model. In this way, the AI / ML model at the first network device 120a can perform a specific task more accurately, especially when the task is related to at least one terminal device.

[0184] In some exemplary embodiments, in order to generate an updated AI / ML model based on at least one AI / ML model, the first network device 120a may also aggregate a fourth AI / ML model and a fine-tuned AI / ML model to generate a fifth AI / ML model as the updated AI / ML model. In this way, the AI / ML model at the first network device 120a can perform specific tasks more accurately, especially when the task is related to at least one terminal device.

[0185] In some exemplary embodiments, the first network device 120a may also send an updated AI / ML model to the second network device and receive a further updated AI / ML model from the second network device. In this way, the AI / ML model at the first network device 120a can perform specific tasks more accurately.

[0186] In some exemplary embodiments, before receiving the AI / ML model from the second network device, the first network device 120a may also send data to the second network device. In this case, the AI / ML model received by the first network device 120a is a fine-tuned AI / ML model based on the data. In this way, the obtained AI / ML model received from the second network device is more accurate for the first network device 120a to perform local tasks.

[0187] In some exemplary embodiments, the first network device may also use at least one of an AI / ML model, a fine-tuned AI / ML model, an updated AI / ML model, or a further updated AI / ML model to perform the task. Alternatively, the first network device 120a may send data related to at least one of a plurality of tasks and stored in an AI / ML database at the first network device 120a to the second network device. In this way, the first network device 120a can perform local tasks using a more accurate AI / ML model.

[0188] In this way, according to method 900, a relatively lightweight custom local AI / ML model can be obtained from the rather large (and “bulky”) global base model at the second network device, thereby reducing the training complexity at the first network device 120a. At the same time, the local AI / ML model at the first network device 120a is more accurate, thus enabling the first network device 120a to perform tasks more accurately.

[0189] FIG. 10 Another flowchart of an exemplary method 1000 implemented at a second network device according to some other embodiments of the present disclosure is shown. For discussion purposes, reference will be made to... FIG. 1B , FIG. 2 and FIG. 4 to 8Method 1000 is described from the perspective of the second network device 120b.

[0190] In box 1010, the second network device 120b is connected from the first network device (e.g., as shown in box 1010). FIG. 2 The first network device 120a shown, or as FIG. 4 The base station 405 shown receives a first request, which instructs the second network device 120b to provide an AI / ML model. In box 1020, the second network device 120b, based on a pre-trained AI / ML model (e.g., such as...), receives a first request. FIG. 1I The pre-trained large model 100I shown, or as... FIG. 4 The global AI model 403 shown generates an AI / ML model. Here, the pre-trained AI / ML model is pre-trained for multiple tasks, including tasks that the first network device will perform using the AI / ML model. In box 1030, the second network device 120b sends the AI / ML model (e.g., as shown in the diagram) to the first network device. FIG. 2 The first AI / ML model 202 shown is illustrated.

[0191] In some exemplary embodiments, the second network device 120b may also receive a second request from the first network device, the second request instructing the second network device to provide a second AI / ML model. In response to the second request, the second network device 120b may generate a second AI / ML model based on a pre-trained AI / ML model and send the second AI / ML model to the first network device. In this way, the second network device 120b may send more than one task-specific AI / ML model to the first network device.

[0192] In some exemplary embodiments, at the second network device 120b, the second request can be received together with the first request, and the second AI / ML model can be sent together with the first AI / ML model. In this way, signaling overhead can be reduced compared to sending the two requests separately.

[0193] In some exemplary embodiments, when transmitting the second AI / ML model along with the first AI / ML model, the second network device 120b may transmit at least one model parameter common to both the first and second AI / ML models, at least one model parameter specific to the first AI / ML model, and at least one model parameter specific to the second AI / ML model. By transmitting the common at least one model parameter once instead of receiving it twice—that is, separately for the first and second AI / ML models—signaling overhead can be reduced.

[0194] In some exemplary embodiments, the second network device 120b may also receive at least one AI / ML model from at least one network device including the first network device, the at least one AI / ML model including an AI / ML model provided by the first network device. Upon receiving the at least one AI / ML model, the second network device 120b may generate an updated AI / ML model based on the at least one AI / ML model and send the updated AI / ML model to the first network device. In this way, the AI / ML model at the first network device can perform local tasks more accurately.

[0195] In some exemplary embodiments, the second network device 120b may also receive at least one AI / ML model from at least one network device including the first network device. The at least one AI / ML model includes an updated AI / ML model provided by the first network device, wherein the updated AI / ML model is generated based on at least one AI / ML model provided by at least one terminal device. Upon receiving the at least one AI / ML model, the second network device 120b may generate an updated AI / ML model based on the at least one AI / ML model and send the updated AI / ML model to the first network device. In this way, the AI / ML model at the first network device can perform a specific task more accurately, especially when the task is related to at least one terminal device.

[0196] In some exemplary embodiments, the AI / ML model sent to the first network device may be a fine-tuned AI / ML model based on data from the first network device. In this way, the AI / ML model at the first network device can perform specific tasks more accurately.

[0197] In some exemplary embodiments, before sending the AI / ML model to the first network device, the second network device 120b may also receive data from the first network device for fine-tuning the AI / ML model, and perform fine-tuning on the AI / ML model based on the received data to obtain a fine-tuned AI / ML pattern. In this way, the AI / ML model at the first network device can be more accurate and "adjusted" to perform a specific task.

[0198] In some exemplary embodiments, the second network device 120b may also receive data from the first network device, which is related to at least one of a plurality of tasks and is stored at the first network device, and store the received data at the second network device 120b. In this way, the global base model at the second network device 120b can be trained with the received data to be more accurate for multiple tasks, and the second network device 120b can generate a more custom AI / ML model specifically for the first network device.

[0199] In this way, according to method 1000, the second network device 120b can provide the first network device with a relatively lightweight custom AI / ML model, rather than a rather large (and "bulky") global base model, thereby reducing the training complexity at the first network device. Simultaneously, the AI / ML model at the first network device is more accurate, thus allowing the first network device to perform tasks more accurately. Furthermore, the second network device 120b can use data received from the first network device to train the global base model for greater accuracy across multiple tasks.

[0200] FIG. 11 Another flowchart illustrating an exemplary method 1100 implemented at a terminal device according to some other embodiments of the present disclosure is shown. For discussion purposes, reference will be made to... FIG. 1B , FIG. 2 , FIG. 5 and FIG. 7 to 8 Method 1100 is described from the perspective of terminal device 110a.

[0201] In box 1110, terminal device 110a receives data from a first network device (e.g., such as...). FIG. 2 The first network device 120a shown, or as FIG. 4 The base station 405 shown receives AI / ML models (e.g., such as...) FIG. 2 The AI / ML model 203 shown is illustrated. In block 1120, terminal device 110a performs fine-tuning on the AI / ML model based on data collected at terminal device 110a to obtain an updated, fine-tuned AI / ML model. In block 1130, terminal device 110a sends the updated AI / ML model to the first network device (e.g., as shown in the diagram). FIG. 2 The updated AI / ML model 204 is shown in the image.

[0202] In this way, according to method 1100, a relatively lightweight custom AI / ML model can be provided to the first network device instead of a rather large (and “bulky” global base model, thereby reducing the training complexity at the first network device. At the same time, the AI / ML model at the first network device is more accurate, so the first network device can perform tasks more accurately (especially tasks related to terminal device 110a).

[0203] FIG. 12 A simplified block diagram of an apparatus 1200 according to some exemplary embodiments of the present disclosure is shown. The apparatus 1200 may be implemented as a device or a chip within a device, and the scope of this application is not limited thereto. The apparatus 1200 may include multiple modules for performing, for example... FIG. 9 The corresponding process in method 900 discussed herein. Apparatus 1200 can be implemented as follows: FIG. 1BThe first network device 120a or a portion thereof shown herein. Reference will be made below. FIG. 1B and 2 describe FIG. 2 .

[0204] like FIG. 4 to 8 As shown, the device 1200 includes a transmitting module 1210, a receiving module 1220, and an obtaining module 1230. The transmitting module 1210 is configured to transmit data, the receiving module 1220 is configured to receive data, and the obtaining module 1230 is configured to obtain data (e.g., obtain an AI / ML model). For example, the transmitting module 1210 is configured to be used in a first network device (e.g., such as...). FIG. 13 The first network device 120a shown in the figure sends data to the second network device (e.g., such as...). FIG. 13 The second network device 120b shown sends a first request (e.g., as shown in the image). FIG. 2 The first request 201 shown in the diagram instructs the second network device to provide a first AI / ML model. The receiving module 1220 is configured to receive the first AI / ML model from the second network device (e.g., as shown in the diagram). FIG. 2 The first AI / ML model 202 shown is illustrated. The obtaining module 1230 is configured to obtain a fine-tuned AI / ML model based on the first AI / ML model.

[0205] In some exemplary embodiments, the obtaining module 1230 may include an execution module for fine-tuning the first AI / ML model based on data from the first network device. In this way, a more accurate local AI / ML model can be obtained.

[0206] In some exemplary embodiments, the apparatus 1200 may further include: a sending module configured to send a second request to the second network device instructing the second network device to provide a second AI / ML model; and a receiving module configured to receive the second AI / ML model from the second network device. In this way, more than one task-specific AI / ML model can be obtained from the second network device.

[0207] In some exemplary embodiments, the second request may be sent together with the first request, and the second AI / ML model may be received together with the first AI / ML model. This reduces signaling overhead compared to sending the two requests separately.

[0208] In some exemplary embodiments, a receiving module configured to receive a second AI / ML model together with a first AI / ML model may include: a receiving module for receiving at least one model parameter common to both the first and second AI / ML models; at least one model parameter specific to the first AI / ML model; and at least one model parameter specific to the second AI / ML model. By receiving at least one common model parameter once instead of twice, i.e., separately for the first and second AI / ML models, signaling overhead can be reduced.

[0209] In some exemplary embodiments, the apparatus 1200 may further include: a sending module configured to send a fine-tuned AI / ML model to a second network device; and a receiving module configured to receive an updated AI / ML model from the second network device. In this way, the AI / ML model at the first network device can perform specific tasks more accurately.

[0210] In some exemplary embodiments, the apparatus 1200 may further include: a transmitting module configured to transmit a fine-tuned AI / ML model to at least one terminal device; a receiving module configured to receive at least one third AI / ML model from at least one terminal device; and a generating module configured to generate an updated AI / ML model based on the at least one third AI / ML model. In this way, the AI / ML model at the first network device can more accurately perform a specific task, particularly when the task is related to at least one terminal device.

[0211] In some exemplary embodiments, at least one terminal device may include multiple terminal devices, at least one third AI / ML model may include multiple AI / ML models, and the generation module may include an aggregation module configured to aggregate multiple AI / ML models to generate a fourth AI / ML model as an updated AI / ML model. In this way, the AI / ML model at the first network device can perform specific tasks more accurately, especially when the task is related to at least one terminal device.

[0212] In some exemplary embodiments, the generation module may further include an aggregation module configured to aggregate the fourth AI / ML model and the fine-tuned AI / ML model to generate a fifth AI / ML model as an updated AI / ML model. In this way, the AI / ML model at the first network device can perform specific tasks more accurately, especially when the task is related to at least one terminal device.

[0213] In some exemplary embodiments, the apparatus 1200 may further include: a sending module configured to send an updated AI / ML model to a second network device; and a receiving module configured to receive a further updated AI / ML model from the second network device. In this way, the AI / ML model at the first network device can perform specific tasks more accurately.

[0214] In some exemplary embodiments, before receiving the AI / ML model from the second network device, the apparatus 1200 may further include a transmission module configured to send data to the second network device, wherein the AI / ML model received by the first network device is a fine-tuned AI / ML model based on the data. In this way, the obtained AI / ML model received from the second network device is more accurate for the first network device.

[0215] In some exemplary embodiments, the apparatus 1200 may further include an execution module configured to perform a task using at least one of an AI / ML model, a fine-tuned AI / ML model, an updated AI / ML model, or a further updated AI / ML model. Alternatively or additionally, the apparatus 1200 may also include a transmission module configured to transmit data associated with at least one of a plurality of tasks and stored in an AI / ML database at the first network device to a second network device. In this way, the first network device can perform localized tasks using a more accurate AI / ML model.

[0216] In this way, a relatively lightweight custom local AI / ML model can be obtained from the rather large (and "bulky") global base model at the second network device, thereby reducing the training complexity at the first network device. At the same time, the local AI / ML model at the first network device is more accurate, so the first network device can perform tasks more accurately.

[0217] FIG. 2 A simplified block diagram of an apparatus 1300 according to some exemplary embodiments of the present disclosure is shown. The apparatus 1300 may be implemented as a device or a chip within a device, and the scope of this application is not limited thereto. The apparatus 1300 may include multiple modules for performing, etc. FIG. 1I The corresponding process in method 1000 discussed herein. Apparatus 1300 can be implemented as follows: FIG. 4 Or the second network device 120b shown in Figure 2, or a portion thereof. Reference will be made below. FIG. 2 , FIG. 14 and FIG. 11 describe FIG. 1B .

[0218] like FIG. 1BAs shown, the device 1300 includes a receiving module 1310, a generating module 1320, and a transmitting module 1330. The receiving module 1310 is configured to receive data, the generating module 1320 is configured to generate data (e.g., generate a custom AI / ML model), and the transmitting module 1330 is configured to transmit data. For example, the receiving module 1310 is configured to be used in a second network device (e.g., such as...). FIG. 2 The second network device 120b shown in the figure receives data from the first network device (e.g., such as...). FIG. 5 The first network device 120a shown receives a first request (e.g., such as...). FIG. 7 to 8 The first request 201 shown in the diagram instructs a second network device to provide an AI / ML model. The generation module 1320 is configured at the second network device to generate a model based on a pre-trained AI / ML model (e.g., ...). FIG. 14 The pre-trained large model shown, or such FIG. 14 The global AI model 403 shown generates an AI / ML model. Here, the pre-trained AI / ML model is pre-trained for multiple tasks, including tasks that the first network device will perform using the AI / ML model. The sending module 1330 is used to send the AI / ML model (e.g., as shown) to the first network device. FIG. 1B The first AI / ML model 202 shown is illustrated.

[0219] In some exemplary embodiments, the apparatus 1300 may further include: a receiving module configured to receive from a first network device a second request instructing a second network device to provide a second AI / ML model; a generating module configured to generate a second AI / ML model based on a pre-trained AI / ML model; and a sending module configured to send the second AI / ML model to the first network device. In this way, the second network device can send more than one task-specific AI / ML model to the first network device.

[0220] In some exemplary embodiments, the second request is received together with the first request, and the second AI / ML model is sent together with the first AI / ML model. This reduces signaling overhead compared to sending the two requests separately.

[0221] In some exemplary embodiments, a generating module configured to transmit a second AI / ML model together with a first AI / ML model may include a transmitting module for transmitting at least one model parameter common to both the first and second AI / ML models; at least one model parameter specific to the first AI / ML model; and at least one model parameter specific to the second AI / ML model. By transmitting the common at least one model parameter once instead of receiving it twice, i.e., separately for the first and second AI / ML models, signaling overhead can be reduced.

[0222] In some exemplary embodiments, the apparatus 1300 may further include: a receiving module configured to receive at least one AI / ML model from at least one network device including a first network device, the at least one AI / ML model including an AI / ML model provided by the first network device; a generating module configured to generate an updated AI / ML model based on the at least one AI / ML model; and a transmitting device configured to transmit the updated AI / ML model to the first network device. In this way, the AI / ML model at the first network device can perform specific tasks more accurately.

[0223] In some exemplary embodiments, the apparatus 1300 may further include: a receiving device configured to receive at least one AI / ML model from at least one network device including a first network device, the at least one AI / ML model including an updated AI / ML model provided by the first network device, wherein the updated AI / ML model is generated based on at least one AI / ML model provided by at least one terminal device; a generating module configured to generate the updated AI / ML model based on the at least one AI / ML model; and a sending module configured to send the updated AI / ML model to the first network device. In this way, the AI / ML model at the first network device can more accurately perform a specific task, especially when the task is related to at least one terminal device.

[0224] In some exemplary embodiments, the AI / ML model sent to the first network device may be a fine-tuned AI / ML model based on data from the first network device. In this way, the AI / ML model at the first network device can perform specific tasks more accurately.

[0225] In some exemplary embodiments, before sending the AI / ML model to the first network device, the apparatus 1300 may further include: a receiving device for receiving data from the first network device for fine-tuning the AI / ML model; and an execution module for fine-tuning the AI / ML model based on the received data to obtain a fine-tuned AI / ML pattern. In this way, the AI / ML model at the first network device can be more accurate and "adjusted" to perform a specific task.

[0226] In some exemplary embodiments, the apparatus 1300 may further include: a receiving module configured to receive data from a first network device that is related to at least one of a plurality of tasks and stored at the first network device; and a storage module configured to store the received data at a second network device. In this way, a global base model at the second network device can be trained using the received data to be more accurate for multiple tasks, and the second network device can generate a more custom AI / ML model specifically for the first network device.

[0227] In this way, a relatively lightweight custom AI / ML model can be provided to the first network device, rather than a rather large (and "bulky") global base model, thereby reducing the training complexity at the first network device. Simultaneously, the AI / ML model at the first network device is more accurate, allowing the first network device to perform tasks more accurately. Furthermore, the second network device can use data received from the first network device to train a global base model for greater accuracy across multiple tasks.

[0228] FIG. 2 A simplified block diagram of an apparatus 1400 according to some exemplary embodiments of the present disclosure is shown. The apparatus 1400 may be implemented as a device or a chip within a device, and the scope of this application is not limited thereto. The apparatus 1400 may include multiple modules for performing, etc. FIG. 5 The corresponding process in method 1100 discussed herein. Apparatus 1400 can be implemented as follows: FIG. 7 to 8 Or the terminal device 110a shown in Figure 2, or a part of the terminal device 110a. Reference will be made below. FIG. 2 , FIG. 4 , FIG. 2 and FIG. 2 describe FIG. 15 .

[0229] like FIG. 1B As shown, the device 1400 includes a receiving module 1410, an execution module 1420, and a sending module 1430. The receiving module 1410 is used to receive data, the execution module 1420 is used to perform operations (e.g., fine-tuning an AI / ML model), and the sending module 1430 is used to send data. For example, the receiving module 1410 is used in a terminal device (e.g., such as...)FIG. 2 , FIG. 4 to 12 , FIG. 4 to 8 and ​ The terminal device 110a shown in the figure receives data from the first network device (e.g., such as...). ​ The first network device 120a shown, or as ​ The base station 405 shown receives AI / ML models (e.g., such as...) ​ The AI / ML model 203 shown is used. The execution module 1420 is used to fine-tune the AI / ML model based on data collected at the terminal device to obtain an updated, fine-tuned AI / ML model. The sending module 1430 is used to send the updated AI / ML model to the first network device (e.g., as shown in the diagram). ​ The updated AI / ML model 204 is shown in the image.

[0230] In this way, the AI / ML model used by the first network device can perform tasks more accurately, especially tasks related to the terminal device.

[0231] ​ A simplified block diagram of a device 1500 suitable for implementing some exemplary embodiments of the present disclosure is shown. The device 1500 can be provided to implement a communication device, such as... ​ and 2 The first network device 120a, the second network device 120b, or the terminal device 110a shown are illustrated. As shown, device 1500 includes one or more processors 1510, one or more memories 1520 coupled to processor 1510, and one or more communication modules 1540 coupled to processor 1510.

[0232] Communication module 1540 is used for bidirectional communication. Communication module 1540 may include a transmitter 1541 for transmitting data and a receiver 1542 for receiving data. Communication module 1540 has at least one antenna to facilitate communication. The communication interface can represent any interface required for communication with other network elements.

[0233] Processor 1510 can be any type suitable for a local technology network and can include one or more of the following: general-purpose computer, special-purpose computer, microprocessor, digital signal processor (DSP), and processor based on a multi-core processor architecture, as non-limiting examples. Device 1500 can have multiple processors, such as application-specific integrated circuit chips that are time-dependent on a clock synchronized with the main processor.

[0234] Memory 1520 may include one or more non-volatile memories and one or more volatile memories. Examples of non-volatile memories include, but are not limited to, read-only memory (ROM) 1524, electrically programmable read-only memory (EPROM), flash memory, hard disk, compact disc (CD), digital video disk (DVD), and other magnetic and / or optical storage. Examples of volatile memories include, but are not limited to, random access memory (RAM) 1522 and other volatile memories that do not persist during power-off periods.

[0235] Computer program 1530 includes computer-executable instructions that are executed by the associated processor 1510. Program 1530 may be stored in ROM 1524. Processor 1510 may perform any suitable actions and processes by loading program 1530 into RAM 1522.

[0236] Embodiments of this disclosure can be implemented via program 1530, enabling device 1500 to execute reference... ​ and ​ Any process discussed in this disclosure. Embodiments of this disclosure may also be implemented in hardware or a combination of software and hardware.

[0237] In some exemplary embodiments, program 1530 may be tangibly included in a computer-readable medium, which may be included in device 1500 (e.g., in memory 1520) or in other storage devices accessible to device 1500. Device 1500 may load program 1530 from the computer-readable medium into RAM 1522 for execution. The computer-readable medium may include any type of tangible non-volatile memory, such as ROM, EPROM, flash memory, hard disk, CD, DVD, etc.

[0238] Generally, the various embodiments of this disclosure can be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects can be implemented in hardware, while others can be implemented in firmware or software, which can be executed by a controller, microprocessor, or other computing device. Although various aspects of the embodiments of this disclosure are shown and described as block diagrams, flowcharts, or using some other graphical representation, it should be understood that the blocks, apparatuses, systems, techniques, or methods described herein can be implemented in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof, as non-limiting examples.

[0239] This disclosure also provides at least one computer program product tangibly stored in a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as those included in a program module, that execute on a device on a target real or virtual processor to perform the functions described above. ​ The described method is 900, 1000, or 1100. Typically, a program module includes routines, programs, libraries, objects, classes, components, data structures, etc., that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of a program module can be combined or split among program modules as needed. The machine-executable instructions of a program module can execute on a local or distributed device. In a distributed device, the program module can reside on both local and remote storage media.

[0240] Program code used to perform the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code implements the functions / operations specified in the flowcharts and / or block diagrams. The program code may be executed entirely on a machine, partially on a machine (as a standalone software package), partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0241] In the context of this disclosure, computer program code or related data may be carried by any suitable carrier to enable a device, apparatus, or processor to perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, etc.

[0242] Computer-readable media can be computer-readable signal media or computer-readable storage media. Computer-readable media can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any suitable combination thereof. Further specific examples of computer-readable storage media will include electrical connections having one or more wires, portable computer floppy disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0243] Furthermore, although operations are shown in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or sequentially, or requiring the execution of all shown operations to achieve the desired result. In some cases, multitasking and parallel processing can be advantageously performed. Similarly, while several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this disclosure, but rather as descriptions of features specific to particular embodiments. Certain features described in the context of individual embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually in multiple embodiments or in any suitable sub-combination.

[0244] Although this disclosure has been described in specific language regarding structural features and / or methodological actions, it should be understood that the disclosure as defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, it discloses the specific features and actions described above as exemplary forms of implementing the claims.

[0245] The following terms may be used in this document.

[0246] LTE Long Term Evolution NR (New Radio) BWP bandwidth part BS base station CA (Carrier Aggregation) CC component carrier CG cell group CSI (Channel State Information) CSI-RS Channel State Information Reference Signal DC dual connectivity DCI downlink control information DL downlink DL-SCH (Downlink Shared Channel) EN-DC E-UTRA NR dual connectivity: MCG uses E-UTRA, SCG uses NR. gNB (Next Generation (or 5G) Base Station) HARQ-ACK Hybrid Automatic Repeat Request Acknowledgement MCG (Master Cell Group) MCS (Modulation and Coding Scheme) MAC-CE Media Access Control - Control Element PBCH (Physical Broadcast Channel) Pcell (primary cell) PDCCH (Physical Downlink Control Channel) PDSCH (Physical Downlink Shared Channel) PRACH (Physical Random Access Channel) PRG (Physical Resource Block Group) PSCell primary SCG cell PSS (Primary Synchronization Signal) PUCCH (Physical Uplink Control Channel) PUSCH (Physical Uplink Shared Channel) RACH (Random Access Channel) RAPID (Random Access Preamble Identity) RB (Resource Block) RE (Resource Element) RRM (Radio Resource Management) RMSI (Remaining System Information) RS reference signal RSRP (Reference Signal Received Power) RRC (Radio Resource Control) SCG (Secondary Cell Group) SFN (System Frame Number) SL sidelink Scell ​​(secondary cell) SPS (semi-persistent scheduling) SR (Scheduling Request) SRI SRS resource indicator SRS sounding reference signal SSS (Secondary Synchronization Signal) SSB (Synchronization Signal Block) SUL (Supplement Uplink) TA (Timing Advance) TAG: Timing Advance Group TUE (Target UE) UCI uplink control information UE (User Equipment) UL uplink UL-SCH Uplink Shared Channel

Claims

1. A method comprising: A first request is sent from the first network device to the second network device, the first request instructing the second network device to provide a first artificial intelligence / machine learning AI / ML model; Receive the first AI / ML model from the second network device; and Based on the first AI / ML model, a fine-tuned AI / ML model is obtained.

2. The method of claim 1, wherein obtaining the fine-tuned AI / ML model comprises: Based on data from the first network device, the first AI / ML model is fine-tuned.

3. The method according to claim 2, further comprising: Send a second request to the second network device, the second request instructing the second network device to provide a second AI / ML model; as well as Receive the second AI / ML model from the second network device.

4. The method of claim 3, wherein the second request is sent together with the first request, and the second AI / ML model is received together with the first AI / ML model.

5. The method of claim 4, wherein receiving the second AI / ML model together with the first AI / ML model includes receiving the following: For at least one model parameter common to both the first AI / ML model and the second AI / ML model; At least one model parameter specific to the first AI / ML model; and At least one model parameter specific to the second AI / ML model.

6. The method according to any one of claims 1 to 5, further comprising: Send the fine-tuned AI / ML model to the second network device; as well as Receive updated AI / ML models from the second network device.

7. The method according to any one of claims 1 to 6, further comprising: The fine-tuned AI / ML model is sent to at least one terminal device; Receive at least one third AI / ML model from the at least one terminal device; as well as An updated AI / ML model is generated based on the at least one third AI / ML model.

8. The method of claim 7, wherein the at least one terminal device comprises a plurality of terminal devices, the at least one third AI / ML model comprises a plurality of AI / ML models, and generating the updated AI / ML model based on the at least one third AI / ML model comprises: The multiple AI / ML models are aggregated to generate a fourth AI / ML model as the updated AI / ML model.

9. The method of claim 8, wherein generating the updated AI / ML model based on the at least one AI / ML model further comprises: The fourth AI / ML model and the fine-tuned AI / ML model are aggregated to generate a fifth AI / ML model as the updated AI / ML model.

10. The method according to any one of claims 7 to 9, further comprising: Send the updated AI / ML model to the second network device; and Receive further updated AI / ML models from the second network device.

11. The method according to claim 2, further comprising: The data is sent to the second network device before the AI / ML model is received from the second network device. The AI / ML model received by the first network device is a fine-tuned AI / ML model based on the data.

12. The method according to any one of claims 1 to 11, further comprising at least one of the following: The task is performed using at least one of the AI / ML model, the fine-tuned AI / ML model, the updated AI / ML model, or the further updated AI / ML model; or Data is sent to the second network device, the data being related to at least one of a plurality of tasks and stored in an AI / ML database at the first network device.

13. A method comprising: A first request is received from the first network device at the second network device, the first request instructing the second network device to provide an artificial intelligence / machine learning (AI / ML) model; The AI / ML model is generated at the second network device based on a pre-trained AI / ML model, wherein the pre-trained AI / ML model is pre-trained for multiple tasks, including tasks that the first network device will perform using the AI / ML model; and The AI / ML model is sent to the first network device.

14. The method of claim 13, further comprising: Receive a second request from the first network device, the second request instructing the second network device to provide a second AI / ML model; The second AI / ML model is generated based on the pre-trained AI / ML model; and The second AI / ML model is sent to the first network device.

15. The method of claim 14, wherein the second request is received together with the first request, and the second AI / ML model is sent together with the first AI / ML model.

16. The method of claim 15, wherein sending the second AI / ML model together with the first AI / ML model includes sending the following: For at least one model parameter common to both the first AI / ML model and the second AI / ML model; At least one model parameter specific to the first AI / ML model; and At least one model parameter specific to the second AI / ML model.

17. The method according to any one of claims 13 to 16, further comprising: Receive at least one AI / ML model from at least one network device including the first network device, the at least one AI / ML model including the AI / ML model provided by the first network device; Based on the at least one AI / ML model, an updated AI / ML model is generated to perform the task; The updated AI / ML model is sent to the first network device.

18. The method according to any one of claims 13 to 17, further comprising: At least one AI / ML model is received from at least one network device including the first network device, the at least one AI / ML model including an updated AI / ML model provided by the first network device, wherein the updated AI / ML model is generated based on at least one AI / ML model provided by at least one terminal device; Based on the at least one AI / ML model, an updated AI / ML model is generated; The updated AI / ML model is sent to the first network device.

19. The method of claim 13, wherein the AI / ML model sent to the first network device is a fine-tuned AI / ML model, the fine-tuned AI / ML model being based on data from the first network device.

20. The method of claim 19, further comprising: Before sending the AI / ML model to the first network device, data for fine-tuning the AI / ML model is received from the first network device; The AI / ML model is fine-tuned based on the received data to obtain the fine-tuned AI / ML mode.

21. The method according to any one of claims 13 to 20, further comprising: Data is received from the first network device, the data being related to at least one of the plurality of tasks and being stored at the first network device; The received data is stored at the second network device.

22. A method comprising: At the terminal device, an artificial intelligence / machine learning (AI / ML) model is received from the first network device. Based on the data collected at the terminal device, the AI / ML model is fine-tuned to obtain an updated AI / ML model; The updated AI / ML model is sent to the first network device.

23. A first network device, comprising: transceiver; The processor is communicatively coupled to the transceiver. The processor is configured as follows: A first request is sent to a second network device via the transceiver, the first request instructing the second network device to provide a first artificial intelligence / machine learning (AI / ML) model; The first AI / ML model is received from the second network device via the transceiver; A fine-tuned AI / ML model is obtained based on the first AI / ML model.

24. A second network device, comprising: transceiver; The processor is communicatively coupled to the transceiver. The processor is configured as follows: The transceiver receives a first request from a first network device, the first request instructing the second network device to provide an artificial intelligence / machine learning (AI / ML) model. The AI / ML model is generated at the second network device based on a pre-trained AI / ML model, wherein the pre-trained AI / ML model is pre-trained for multiple tasks, including tasks that the first network device will perform using the AI / ML model; and The AI / ML model is sent to the first network device via the transceiver.

25. A terminal device, comprising: transceiver; The processor is communicatively coupled to the transceiver. The processor is configured as follows: Receive artificial intelligence / machine learning (AI / ML) models from a first network device via the transceiver; Based on the data collected at the terminal device, the AI / ML model is fine-tuned to obtain an updated AI / ML model; The updated AI / ML model is sent to the first network device via the transceiver.

26. A non-transient computer-readable medium comprising a computer program stored thereon, which, when executed on at least one processor, causes the at least one processor to perform the method according to any one of claims 1 to 22.

27. A chip comprising at least one processing circuit configured to perform the method according to any one of claims 1 to 22.

28. A computer program product tangibly stored on a computer-readable medium and comprising computer-executable instructions that, when executed, cause a device to perform the method according to any one of claims 1 to 22.

Images