Transfer learning in vertical domains using the application enablement layer

Abstract

A method is provided that includes sending a machine learning (ML) model retrieval request to a server supporting a source vertical application layer (VAL) domain. The method includes receiving a ML model retrieval response from the server based on the ML model retrieval request, the ML model retrieval response including at least one candidate ML model built using training data from the source VAL domain. And the method includes performing a transfer learning operation in which a source ML model selected from the at least one candidate ML model is tuned using training data from a target VAL domain to build a target ML model.

Description

TRANSFER LEARNING IN VERTICAL DOMAINS USING THE APPLICATION ENABLEMENT LAYERTECHNOLOGICAL FIELD

[0001] The present disclosure relates generally to telecommunications and, in particular, to support for vertical applications over a telecommunications network.BACKGROUND

[0002] A telecommunications system can be seen as a facility that enables communication sessions between two or more entities such as user terminals, base stations and / or other nodes by providing carriers between the various entities involved in the communications path. A telecommunications system can be provided for example by means of a communication network and one or more compatible communication devices. The communication sessions may comprise, for example, communication of data for carrying communications such as voice, video, electronic mail (email), text message, multimedia and / or content data and so on. Non-limiting examples of services provided comprise two-way or multi-way calls, data communication or multimedia services and access to a data network system, such as the Internet.

[0003] In a wireless telecommunications system at least a part of a communication session between at least two stations occurs over a wireless link. Examples of wireless systems comprise public land mobile networks (PLMN), satellite based communication systems and different wireless local networks, for example wireless local area networks (WLAN). Some wireless systems can be divided into cells, and are therefore often referred to as cellular systems.

[0004] A user can access the telecommunications system by means of an appropriate communication device or terminal. A communication device of a user may be referred to as user equipment (UE) or user device. A communication device is provided with an appropriate signal receiving and transmitting apparatus for enabling communications, for example enabling access to a communication network or communications directly with other users. The communication device may access a carrier provided by a station, for example a base station of a cell, and transmit and / or receive communications on the carrier.

[0005] The telecommunications system and associated devices typically operate in accordance with a given standard or specification which sets out what the various entities associated with the system are permitted to do and how that should be achieved. Communication protocols and / or parameters which shall be used for the connection are also typically defined. One example of a telecommunications system is the Universal Mobile Telecommunications System (UMTS). Other examples of telecommunications systems are Long-Term Evolution (LTE), LTE Advanced and the so-called 5G or New Radio (NR) networks. NR is being standardized by the 3rd Generation Partnership Project (3GPP).BRIEF SUMMARY

[0006] Example implementations of the present disclosure are directed to telecommunications and, in particular, to support for vertical applications over a telecommunications network. The present disclosure includes, without limitation, the following example implementations.

[0007] Some example implementations provide an apparatus comprising: at least one memory configured to store instructions; and at least one processing circuitry configured to access the at least one memory, and execute the instructions to cause the apparatus to at least: send a machine learning (ML) model retrieval request to a server supporting a source vertical application layer (VAL) domain; receive a ML model retrieval response from the server based on the ML model retrieval request, the ML model retrieval response including at least one candidate ML model built using training data from the source VAL domain; and perform a transfer learning operation in which a source ML model selected from the at least one candidate ML model is tuned using training data from a target VAL domain to build a target ML model.

[0008] Some example implementations provide an apparatus comprising: means for sending a machine learning (ML) model retrieval request to a server supporting a source vertical application layer (VAL) domain; means for receiving a ML model retrieval response from the server based on the ML model retrieval request, the ML model retrieval response including at least one candidate ML model built using training data from the source VAL domain; and means for performing a transfer learning operation in which a source ML modelselected from the at least one candidate ML model is tuned using training data from a target VAL domain to build a target ML model.

[0009] Some example implementations provide a method comprising: sending a machine learning (ML) model retrieval request to a server supporting a source vertical application layer (VAL) domain; receiving a ML model retrieval response from the server based on the ML model retrieval request, the ML model retrieval response including at least one candidate ML model built using training data from the source VAL domain; and performing a transfer learning operation in which a source ML model selected from the at least one candidate ML model is tuned using training data from a target VAL domain to build a target ML model.

[0010] Some example implementations provide a computer-readable storage medium that is non-transitory and has instructions stored therein that, in response to execution by at least one processing circuitry, causes an apparatus to at least: send a machine learning (ML) model retrieval request to a server supporting a source vertical application layer (VAL) domain; receive a ML model retrieval response from the server based on the ML model retrieval request, the ML model retrieval response including at least one candidate ML model built using training data from the source VAL domain; and perform a transfer learning operation in which a source ML model selected from the at least one candidate ML model is tuned using training data from a target VAL domain to build a target ML model.

[0011] Some example implementations provide an apparatus comprising: at least one memory configured to store instructions; and at least one processing circuitry configured to access the at least one memory, and execute the instructions to cause the apparatus to at least: receive a machine learning (ML) model retrieval request from a server supporting a target vertical application layer (VAL) domain; retrieve at least one candidate ML model based on the ML model retrieval request, the at least one candidate ML model built using training data from a source VAL domain; and send a ML model retrieval response to the server that includes the at least one candidate ML model, the ML model retrieval response sent to enable a transfer learning operation in which a source ML model selected from the at least one candidate ML model is tuned using training data from the target VAL domain to build a target ML model.

[0012] Some example implementations provide an apparatus comprising: means for receiving a machine learning (ML) model retrieval request from a server supporting a target vertical application layer (VAL) domain; means for retrieving at least one candidate ML modelbased on the ML model retrieval request, the at least one candidate ML model built using training data from a source VAL domain; and means for sending a ML model retrieval response to the server that includes the at least one candidate ML model, the ML model retrieval response sent to enable a transfer learning operation in which a source ML model selected from the at least one candidate ML model is tuned using training data from the target VAL domain to build a target ML model.

[0013] Some example implementations provide a method comprising: receiving a machine learning (ML) model retrieval request from a server supporting a target vertical application layer (VAL) domain; retrieving at least one candidate ML model based on the ML model retrieval request, the at least one candidate ML model built using training data from a source VAL domain; and sending a ML model retrieval response to the server that includes the at least one candidate ML model, the ML model retrieval response sent to enable a transfer learning operation in which a source ML model selected from the at least one candidate ML model is tuned using training data from the target VAL domain to build a target ML model.

[0014] Some example implementations provide a computer-readable storage medium that is non-transitory and has instructions stored therein that, in response to execution by at least one processing circuitry, causes an apparatus to at least: receive a machine learning (ML) model retrieval request from a server supporting a target vertical application layer (VAL) domain; retrieve at least one candidate ML model based on the ML model retrieval request, the at least one candidate ML model built using training data from a source VAL domain; and send a ML model retrieval response to the server that includes the at least one candidate ML model, the ML model retrieval response sent to enable a transfer learning operation in which a source ML model selected from the at least one candidate ML model is tuned using training data from the target VAL domain to build a target ML model.

[0015] These and other features, aspects, and advantages of the present disclosure will be apparent from a reading of the following detailed description together with the accompanying figures, which are briefly described below. The present disclosure includes any combination of two, three, four or more features or elements set forth in this disclosure, regardless of whether such features or elements are expressly combined or otherwise recited in a specific example implementation described herein. This disclosure is intended to be read holistically such that any separable features or elements of the disclosure, in any of its aspects and exampleimplementations, should be viewed as combinable unless the context of the disclosure clearly dictates otherwise.

[0016] It will therefore be appreciated that this Brief Summary is provided merely for purposes of summarizing some example implementations so as to provide a basic understanding of some aspects of the disclosure. Accordingly, it will be appreciated that the above described example implementations are merely examples and should not be construed to narrow the scope or spirit of the disclosure in any way. Other example implementations, aspects and advantages will become apparent from the following detailed description taken in conjunction with the accompanying figures which illustrate, by way of example, the principles of some described example implementations.BRIEF DESCRIPTION OF THE FIGURE(S)

[0017] Having thus described example implementations of the disclosure in general terms, reference will now be made to the accompanying figures, which are not necessarily drawn to scale, and wherein:

[0018] FIG. 1 illustrates a telecommunications system that includes one or more public land mobile networks (PLMNs) coupled to one or more external data networks, according to some example implementations of the present disclosure;

[0019] FIG. 2 illustrates a deployment of a PLMN, according to some example implementations;

[0020] FIG. 3 more particularly depicts aspects of the deployment of FIG. 2, according to some example implementations;

[0021] FIG. 4 illustrates an on-network functional model for ADAE, according to some example implementations;

[0022] FIG. 5 illustrates an ADAE-based deployment including two vertical domains, and in which one or more artificial intelligence (Al) / machine learning (ML) (AIML) services may be executed, according to some example implementations;

[0023] FIGS. 6Aand 6B illustrate a signaling chart of transfer learning between two vertical domains in an ADAE deployment, according to some example implementations;

[0024] FIGS. 7Aand 7B are flowcharts illustrating various steps in a method according to various example implementations;

[0025] FIGS. 8 A and 8B are flowcharts illustrating various steps in a method according to various example implementations; and

[0026] FIG. 9 illustrates an apparatus according to some example implementations.DETAILED DESCRIPTION

[0027] Some implementations of the present disclosure will now be described more fully hereinafter with reference to the accompanying figures, in which some, but not all implementations of the disclosure are shown. Indeed, various implementations of the disclosure may be embodied in many different forms and should not be construed as limited to the implementations set forth herein; rather, these example implementations are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art. Like reference numerals refer to like elements throughout.

[0028] Unless specified otherwise or clear from context, references to first, second or the like should not be construed to imply a particular order. A feature described as being above another feature (unless specified otherwise or clear from context) may instead be below, and vice versa; and similarly, features described as being to the left of another feature else may instead be to the right, and vice versa. Also, while reference may be made herein to quantitative measures, values, geometric relationships or the like, unless otherwise stated, any one or more if not all of these may be absolute or approximate to account for acceptable variations that may occur, such as those due to engineering tolerances or the like.

[0029] As used herein, unless specified otherwise or clear from context, the “or” of a set of operands is the “inclusive or” and thereby true if and only if one or more of the operands is true, as opposed to the “exclusive or” which is false when all of the operands are true. Thus, for example, “[A] or [B]” is true if [A] is true, or if [B] is true, or if both [A] and [B] are true. Further, the articles “a” and “an” mean “one or more,” unless specified otherwise or clear from context to be directed to a singular form. Furthermore, it should be understood that unless otherwise specified, the terms “data,” “content,” “digital content,” “information,” and similar terms may be at times used interchangeably. The term “network” may refer to a group of interconnected computers including clients and servers; and within a network, these computers may be interconnected directly or indirectly by various means including via one or more switches, routers, gateways, access points or the like.

[0030] Reference may be made herein to terms specific to a particular system, architecture or the like, but it should be understood that example implementations of the present disclosure may be equally applicable to any of a number of systems, architectures and the like. For example, reference may be made to 3 GPP technologies such as Global System for Mobile Communications (GSM), UMTS, LTE, LTE Advanced, 5GNR, 5G Advanced and 6G; however, it should be understood that example implementations of the present disclosure may be equally applicable to non-3GPP technologies such as IEEE 802, Bluetooth and Bluetooth Low Energy.

[0031] Further, as used in this application, the term “circuitry” may refer to one or more or all of the following: (a) hardware-only circuit implementations (such as implementations in only analog and / or digital circuitry); (b) combinations of hardware circuits and software, such as (as applicable): (i) a combination of analog and / or digital hardware circuit(s) with software / firmware and (ii) any portions of hardware processor(s) with software (including digital signal processor(s)), software, and memory(ies) that work together to cause an apparatus, such as a mobile phone or server, to perform various functions); or (c) hardware circuit(s) and / or processor(s), such as a microprocessor(s) or a portion of a microprocessor s), that requires software (e.g., firmware) for operation, but the software may not be present when it is not needed for operation.

[0032] The above definition of circuitry applies to all uses of this term in this application, including in any claims. As a further example, as used in this application, the term circuitry also covers an implementation of merely a hardware circuit or processor (or multiple processors) or portion of a hardware circuit or processor and its (or their) accompanying software and / or firmware. The term circuitry also covers, for example and if applicable to the particular claim element, a baseband integrated circuit or processor integrated circuit for a mobile device or a similar integrated circuit in server, a cellular network device, or other computing or network device.

[0033] FIG. 1 illustrates a telecommunications system 100 according to various example implementations of the present disclosure. The telecommunications system generally includes one or more telecommunications networks. As shown, for example, the system includes one or more public land mobile networks (PLMNs) 102 coupled to one or more other external data networks 104 - notably including a wide area network (WAN) such as the Internet. Each ofthe PLMNs includes a core network (CN) 106 backbone such as the Evolved Packet Core (EPC) of LTE, the 5G core network (5GC) or the like; and each of the core networks and the Internet are coupled to one or more radio access networks (RANs) 108, air interfaces or the like that implement one or more radio access technologies (RATs). As used herein, a “network device” refers to any suitable device at a network side of a telecommunications network. Examples of suitable network devices are described in greater detail below.

[0034] In addition, the system includes one or more radio units that may be varyingly known as user equipment (UE) 110, terminal device, terminal equipment, mobile station or the like. The UE is generally a device configured to communicate with a network device or a further UE in a telecommunication network. The UE may be a portable computer (e.g., laptop, notebook, tablet computer), mobile phone (e.g., cell phone, smartphone), wearable computer (e.g., smartwatch), or the like. In other examples, the UE may be an Internet of Things (loT) device, an industrial loT (IIoT device), a vehicle equipped with a vehicle-to- everything (V2X) communication technology, or the like. In some examples, as referenced by 3 GPP, the UE may be a narrowband loT (NB-IoT) device, an enhanced machine-type communication (eMTC) device, a reduced capability (RedCap) device, an ambient loT device, or the like.

[0035] In operation, these UEs may be configured to connect to one or more of the RANs 108 according to their particular radio access technologies to thereby access a particular CN 106 of a PLMN 102, or to access one or more of the external data networks 104 (e.g., the Internet). The external data network may be configured to provide Internet access, operator services, 3rd party services, etc. For example, the International Telecommunication Union (ITU) has classified 5G mobile network services into three categories: enhanced mobile broadband (eMBB), ultra-reliable and low-latency communications (URLLC), and massive machine type communications (mMTC) or massive internet of things (MIoT).

[0036] Examples of radio access technologies include 3 GPP radio access technologies such as GSM, UMTS, LTE, LTE Advanced, 5GNR, 5G Advanced, and 6G. Other examples of radio access technologies include IEEE 802 technologies such as IEEE 802.11 (Wi-Fi), IEEE 802.15 (including 802.15.1 (WPAN / Bluetooth), 802.15.4 (Zigbee) and 802.15.6 (WBAN)), Bluetooth, Bluetooth Low Energy (BLE), ultra wideband (UWB), and the like. Generally, a radio access technology may refer to any 2G, 3G, 4G, 5G, 6G or higher generation mobile communication technology and their different versions, as well as to any other wireless radioaccess technology that may be arranged to interwork with such a mobile communication technology to provide access to the CN 106 of a mobile network operator (MNO).

[0037] In various examples, a RAN 108 may be configured as one or more macrocells, microcells, picocells, femtocells or the like. The RAN may generally include one or more radio access nodes that are configured to interact with UEs 110. In various examples, a radio access node may be referred to as a base station (BS), access point (AP), base transceiver station (BTS), Node B (NB), evolved NB (eNB), macro BS, NB (MNB) or eNB (MeNB), home BS, NB (HNB) or eNB (HeNB), next generation NB (gNB), enhanced gNB (en-gNB), next generation eNB (ng-eNB), or the like. The RAN may include some type of network controlling / governing entity responsible for control of the radio access nodes. The network controlling / governing entity and radio access node may be separate or integrated into a single apparatus. The network controlling / governing entity may include processing circuity configured to carry out various management functions, etc. The processing circuity may be associated with a memory, computer-readable storage medium or database for maintaining information required in the management functions.

[0038] ARAN 108 may be centralized or distributed. In various examples, components of a RAN may be interconnected by Ethernet, Gigabit Ethernet, Asynchronous Transfer Mode (ATM), optical fiber, dark fiber, passive wavelength division multiplexing (WDM), WDM passive optical network (WDM-PON), optical transport network (OTN), time sensitive networking (TSN) and / or any other data link layer network, possibly including radio links. The RAN may be connected to a CN 106 through one or more gateways, network functions or the like.

[0039] As will be appreciated, a PLMN 102 may be deployed in a number of different manners. In a 4GLTE deployment, the EPC is the CN 106, and the evolved UMTS terrestrial radio access network (E-UTRAN) is the RAN 108; and the E-UTRAN includes one or more eNBs (radio access nodes) configured to connect UEs 110 to the E-UTRAN to thereby access the EPC. FIG. 2 illustrates a deployment 200, such as a 5G or 6G deployment. As shown, the 5GC 202 is the CN, and the next generation (NG) radio access network (NG-RAN) 204 is the RAN; and the NG-RAN includes one or more gNBs 206 (radio access nodes) configured to connect UEs 208 to the NG-RAN to thereby access the 5GC. The term ‘gNB’ in 5G may correspond to the eNB in 4G LTE.

[0040] Some 4GLTE and 5G deployments are considered standalone (SA) deployments. Other deployments combine 4GLTE and 5 G technologies, and are referred to as non- standalone (NSA) deployments. In some deployments, the E-UTRAN includes one or more ng-eNBs that are configured to communicate with the 5GC, and that may also be configured to communicate with one or more gNBs. Similarly, in another deployment, the NG-RAN may include one or more en-gNBs that are configured to communicate with the EPC, and that may also be configured to communicate with one or more eNBs. In various instances, a single UE 110, a dual-mode or multimode UE, may support multiple (two or more) RANs — thereby being configured to connect to multiple RANs, such as 4G LTE and 5G.

[0041] FIG. 3 more particularly depicts aspects of the deployment 200 for a MNO, according to some example implementations. As shown, the deployment includes the 5GC 202, and NG-RAN 204 with one or more gNBs 206 configured to connect UEs 208 to the NG- RAN to thereby access the 5GC. The 5GC may include a number of network functions (NFs) divided between the control plane and the user plane. In particular, the 5GC may include, for example, an access and mobility management function (AMF) 302, a session management function (SMF) 304, a user plane function (UPF) 306, a network exposure function (NEF) 308, a network data analytics function (NWDAF) 310, and / or an application function (AF) 312. Other examples of suitable NFs include a network repository function (NRF), a network slice selection function (NSSF), a policy control function (PCF), a unified data management (UDM), or the like. Also shown is a server hosting an application, referred to as application server (AS) 314.

[0042] In the control plane, the AMF 302 is configured to provide UE-based authentication, authorization, mobility management, etc. The SMF 304 is configured to provide various functionality including session management (SM), UE Internet Protocol (IP) address allocation and management, selection and control of UPF(s) 306, control part of policy enforcement and Quality of Service (QoS), lawful intercept, termination of SM parts of NAS messages, Downlink Data Notification (DNN), roaming functionality, handle local enforcement to apply QoS for Service Level Agreements (SLAs), charging data collection and charging interface, etc. If the UE 208 has multiple sessions, different SMFs may be allocated to each session to manage them individually and possibly provide different functionalities per session.

[0043] The UPF 306 supports various user plane operations and functionalities, such as packet routing and forwarding, traffic handling (e.g., QoS enforcement), an anchor point for intra-RAT / inter-RAT mobility (when applicable), packet inspection and policy rule enforcement, lawful intercept (UP collection), traffic accounting and reporting, etc. The UPF is the point of interconnect between the 5GC and at least one external data network (DN) 316 (i.e., point of ingress or egress for a DN), and routes packets to and from the DN. The DN may be configured to provide Internet access, operator services, 3rd party services, etc.

[0044] The NWDAF 310 collects and analyzes network data to provide insights into the performance, optimization, and overall health of the 5GC 202. This information may be used for network management, optimization, and decision-making processes to enhance the network’s efficiency.

[0045] The AF 312 may interact with the 5GC 202 to enable the deployment of specific services and applications. The AF communicates with other NFs to request and manage network resources, ensuring that the network adapts to the requirements of different applications and services. The NEF 308 allows authorized third-party applications and services to access specific network functions and services in a controlled manner. The NEF enables the exposure of network capabilities to external entities, fostering innovation and the development of new services.

[0046] In some deployments, such as deployment 200, operations of the gNB 206 or other radio access node may be carried out, at least partly, in a central / centralized unit (CU), such as a server, host or node, operationally coupled to a distributed unit (DU), such as a radio head / node. It is also possible that node operations may be distributed among a plurality of servers, hosts or nodes. It should also be understood that the distribution of work between 5GC 202 (or other CN) operations and gNB (or other radio access node) operations may vary depending on implementation.

[0047] A 5G network architecture may be based on a so-called CU-DU split. One gNB- CU (central node) may control one or more gNB-DUs. The gNB-CU may control a plurality of spatially separated gNB-DUs, acting at least as transmit / receive (Tx / Rx) nodes. In some example implementations, however, the gNB-DUs (also called DU) may include, for example, a radio link control (RLC), medium access control (MAC) layer and a physical (PHY) layer, whereas the gNB-CU (also called a CU) may include the layers above the RLC layer, such as apacket data convergence protocol (PDCP) layer, a radio resource control (RRC), and an internet protocol (IP) layer. Other functional splits are also possible. It is considered that a skilled person is familiar with the open systems interconnection (OSI) model and the functionalities within each layer.

[0048] In some example implementations, the server or CU may generate a virtual network through which the server communicates with the radio node. In general, virtual networking may involve a process of combining hardware and software network resources and network functionality into a single, software-based administrative entity, a virtual network. Such virtual network may provide flexible distribution of operations between the server and the radio head / node. In practice, any digital signal processing task may be performed in either the CU or the DU, and the boundary where the responsibility is shifted between the CU and the DU may be selected according to implementation.

[0049] In 3 GPP, an application enablement layer is provided to support third-party application developers and vertical-specific application providers. In the application enablement layer, the service enabler architecture layer (SEAU) supports communication and functionality between vertical applications (e.g., V2X applications) and the underlying network. The SEAE architecture includes a common set of services (e.g. group management, location management) and reference points. The SEAL offers its services to a vertical application layer (VAL), and the architecture includes functional entities on a UE 110 (referred to at times as a VAL UE) and a server that are grouped into SEAL client(s) and SEAL server(s).

[0050] In particular, the SEAL architecture includes one or more VAL clients that provide client-side functionalities corresponding to one or more vertical applications, and one or more VAL servers that provides server-side functionalities of the corresponding vertical application(s). Similarly, one or more SEAL clients provide client-side functionalities corresponding to one or more SEAL services, and one or more SEAL servers provide serverside functionalities of the corresponding SEAL service(s). The VAL client(s) communicates with the VAL server(s) across a 3 GPP network system over a VAL-UU reference point, and the SEAL client(s) communicates with the SEAL server(s) across the 3 GPP network system over a SEAL-UU reference point.

[0051] The application data analytics enabler (ADAE) is a SEAL service that offers application data analytics capabilities. ADAE considers vertical-specific applications and edge applications as the major consumers of 3 GPP-provided data analytics services, and supports a number of different analytics. In particular, for example, ADAE supports application performance analytics, slice-specific application performance analytics, UE-to-UE application performance analytics, location accuracy analytics, service application programming interface (API) analytics, slice usage pattern analytics, and edge load analytics.

[0052] FIG. 4 illustrates an on-network functional model 400 for ADAE, according to some example implementations of the present disclosure. The ADAE is a service of SEAL, which offers its services to the VAL; and as shown, the VAL in a vertical (VAL) domain 402 includes VAL client(s) 404 (on a VAL UE 406) and VAL server(s) 408. Similarly, the ADAE includes one or more ADAE clients 410 (on the VAL UE) and one or more ADAE servers 412. The ADAE client and the ADAE server support one or more VAL applications. The ADAE client and ADAE server provide respectively client-side functionalities and server-side functionalities.

[0053] The ADAE client 410 communicates with the ADAE server 412 across the 3GPP network system 414 over an ADAE -UU reference point. The ADAE client provides support for ADAE functions to the VAL client(s) 404 over an ADAE-C reference point. The VAL server(s) 408 communicates with the ADAE server over an ADAE-S reference point. The ADAE server, acting as an AF 312, may communicate with 5GC functions over the N33 reference point to the NEF 308 and N6 reference point to the UPF 306), and operations, administration and management (0AM) over an ADAE-OAM interface.

[0054] As also shown, the ADAE server 412 is a component part of an ADAE internal architecture that operates in a DN 316 or edge DN. The ADAE internal architecture may also include a coordination function, which in some examples may be referred to as an application layer - data collection and coordination function (A-DCCF) 414. Additionally or alternatively, the ADAE internal architecture may include a repository function, which in some examples may be referred to as an application layer - analytics and data repository function (A- ADRF) 416. The A-DCCF coordinates collection and distribution of requested data, and the A- ADRF stores historical data and / or analytics, i.e., data and / or analytics related to past time period that has been obtained by an ADAE server 412 (via AADRF-1) or other NFs / NWDAF 310. TheA-DCCF may be used to fetch data or put data into an application-level entity, such as the A- ADRF or one or more data sources 418. The A-DCCF may coordinate the collection and distribution of data requested by an ADAE server (over ADCCF-1, ADAE-X), although the ADAE server may also directly interact with the data sources via ADAE- Y.

[0055] As a SEAL service that offers application data analytics capabilities, the ADAE may be enhanced to support machine learning (ML) operations, like (re)training or performing inference. One of the main challenges for training ML models is to obtain enough representative data (training data) to accurately capture the input-output relationship of the task of interest. Normally, complex dynamics in mobile networks can only be captured using complex ML models (e.g., deep neural networks) that require a lot of training data. But a sufficient amount of training data is not always available right away, such as when a nonpublic network (NPN) is installed in a new factory branch, or an loT network is deployed in a new set of environmental conditions.

[0056] It has been stated that the 5G system should support at least three types of artificial intelligence (Al) / ML operations, namely, AI / ML operation splitting between AI / ML endpoints, distributed / federated learning over the 5G system, and AI / ML model / data distribution and sharing over the 5G system.

[0057] For AI / ML model / data distribution and sharing over the 5G system, it may be expected to let AI / ML consumers select an ML model out of a set of candidate models available in a network endpoint / repository) to better adapt to task and environment variations. In practical deployments involving AI / ML operations, however, the number of available candidate models is not that large to cover all possible environmental conditions in a network. Out-of-the-box usage of a candidate model may imply a lack of accuracy / precision during an inference stage unless a model fine tuning stage is present to ensure that the baseline model is quickly adapted to current conditions using a lesser amount of data (than the one originally used for training of the baseline model).

[0058] Example implementations of the present disclosure therefore provide a solution that exploits the concept of ML model distribution and leverages reuse of “learnings” from a source domain into a target domain. According to some example implementations, a baseline ML model (e.g., a deep neural network) can be quickly fine-tuned (using a rather small dataset size) to adapt to new network / environmental conditions. Additionally or alternatively, thetarget domain may decide the portion of the ML model that can be reused (and therefore the part that is fine-tuned) using information about data collected locally in the target domain (e.g., the number of samples collected), which can be considered private information in the target information.

[0059] Some example implementations enable verticals to use transfer learning (TL) to allow a faster adoption of ML models as well as a better / faster adaptation to new environment conditions solving one of the main challenges in ML that is the lack of data for training models from scratch. In this regard, TL is a technique that allows quickly exploiting the use of an already-trained ML model (trained in a so-called source domain) in a new target domain. TL can be applied to areas where the knowledge of an already-trained ML model can be utilized or adapted to a different (but related) domain. Instead of “learning” from scratch in the target domain, the learning may be carried forward with patterns / features already learned while solving a “related” problem in the source domain.

[0060] According to some example implementations, a NF / entity may be enhanced with one or more Al / ML (AIML) services to train and fine tune ML models provided enough historical data is available. These AIML services may be present in one or more entities at the same time, such as in an ADAE-based deployment as shown in FIG. 5.

[0061] FIG. 5 illustrates an ADAE-based deployment 500 including two vertical domains, and in which AIML services(s) may be executed, according to some example implementations. The two vertical domains may be similar or different in nature. For example, one of the vertical domains may relate to V2X, while the other vertical domain relates to an industrial scenario. In other examples, both domains may relate to V2X, an industrial scenario, or another scenario. Additionally, the vertical domains may be administrated by the same VAL service provider; or the vertical domains may be administered by different VAL service providers, if a business relationship (to share local “learnings”) exists between the different VAL service providers.

[0062] In some more particular examples, as shown, the ADAE-based deployment 500 includes a first VAL domain (VAL domain #1 or source VAL domain) 402A and a second VAL domain (VAL domain #2 or target VAL domain) 402B, according to some example implementations. In the deployment, each of the vertical domains includes a respective first VAL UE (VAL UE #1) 406A and second VAL UE (VAL UE #2) 406B. The VAL UEs includerespective first and second VAL client(s) 404A, 404B that communicate with respective first and second VAL servers 408A, 408B, and the VAL UEs include respective first and second ADAE clients 410A, 410B that communicate with respective first and second ADAE servers 412A, 412B

[0063] A number of the entities in the ADAE-based deployment 500 provide one or more AIML services 502. In some examples, the first and second ADAE clients 410A, 410B that are enhanced with AIML services may be referred to as first and second AIML enablement clients. Similarly, in some examples, the first and second ADAE servers 412A, 412B that are enhanced with AIML services may be referred to as first and second AIML enablement servers. In other examples, either or both of the clients or servers may be more generally referred to as clients or servers (or referred by other names) that are enhanced with AIML services, such as (re)training ML models and performing inference.

[0064] In the ADAE-based deployment 500, once an ML model is generated by an AIML producer in the source VAL domain 402A, such as by a first AIML enablement server (the first ADAE server 412A with AIML services 502), a number of functionalities may be enabled.The first AIML enablement server may itself perform local inference tasks using the ML model, or the ML model may be distributed to several entities within the same VAL domain (e.g., first VAL server 408A or first AIML enablement client (first ADAE client 410A with AIML services). If a business relationship is in place, the ML model(s) may be distributed to entities outside the source VAL domain, such as to the second VAL server 408B or a second AIML enablement server (the second ADAE server 412B enhanced to provide AIML services). The ML model may also be stored in a local repository for future use, such as by A- ADRF 416 or another repository function.

[0065] Whenever an entity in the target VAL domain 402B (e.g., second VAL server 408B) wants to solve a task using a ML model, enough data to train a tailored ML model from scratch is not yet available, the second VAL server may use TL capabilities, which may be based on a business agreement with another vertical domain. In this regard, the second VAL server may retrieve and fine tune a ML model (out-of-a set of candidate ML models) from the source VAL domain 402A. In some examples, retrieval of the ML model may be based on one or more filtering criteria. Examples of suitable filtering criteria include type of task to be performed by the ML model (e.g., regression / prediction, classification, clustering), number ofinputs, number of output(s), data types in the inputs and output(s), performance metrics of the ML model for training / validation / test sets, or the like.

[0066] If candidate ML model(s) that match the filtering criteria are available in a repository for the source VAL domain 402A, the second VAL server 408B may retrieve the candidate ML model(s) and use AIML services 502 instantiated on it to perform a TL operation in which a particular source ML model of the set of candidate ML model(s) is finetuned using a lesser amount of data collected in the target VAL domain 402B. This TL operation may follow a policy that depends on a number of samples of locally-collected data, and complexity of the ML model. In this regard, the more the data locally collected, the more components of the source ML model that can be adjusted / fine-tuned). Also, the higher the complexity of the source ML model, the greater the amount of data that may be needed for fine tuning.

[0067] Some example implementations of the present disclosure therefore provide a first AIML enablement server (first ADAE server 412A with AIML services 502) supporting a source VAL domain 402A, and a second AIML enablement server (second ADAE server 412B with AIML services) supporting a target VAL domain 402B. In some examples, the source VAL domain supported by the first AIML enablement server is included in a trust domain of the target VAL domain. In particular, for example, the source and target VAL domains may belong to the same VAL service provider, or belong to respective VAL service providers that have a business relationship in place.

[0068] In some examples, the second AIML enablement server (second ADAE server 412B with AIML services 502) may be configured to make a determination that a repository function (e.g., an A-ADRF) (not separately shown in FIG. 5) supporting the target VAL domain 402B is unable to find at least one matching ML model based on one or more filtering criteria.

[0069] The second AIML enablement server (second ADAE server 412B with AIML services 502) may be configured to send a (first) ML model retrieval request to the first AIML enablement server (first ADAE server 412A with AIML services) supporting the source VAL domain 402A, such as based on the determination by the second AIML enablement server. In some examples, the ML model retrieval request includes the same or different one or more filtering criteria. The first AIML enablement server may be configured to receive the (first)ML model retrieval request, and retrieve at least one candidate ML model (built using training data from the source VAL domain) based on the ML model retrieval request. In this regard, the first AIML enablement server may be configured to send a (second) ML model retrieval request to A-ADRF 416 or other repository function supporting the source VAL domain, and receive a (first) ML model retrieval response including the candidate ML model(s) from the A- ADRF based on the ML model retrieval request.

[0070] In some examples, the candidate ML model(s) may be found or retrieved by the first AIML enablement server (first ADAE server 412A with AIML services 502) based on the one or more filtering criteria.

[0071] In some examples, the candidate ML model(s) may be built using training data from the source VAL domain that is at least one of normalized, anonymized or obfuscated to protect sensitive information of the source VAL domain from the target VAL domain 402B. In this regard, the source VAL domain 402A may incorporate an anonymization / obfuscating stage of both the variables involved in the ML model training and inference. More particularly, for example, the variables may be normalized to a specific range (e.g., the so-called normalization range between 0 and 1, so the scale of the ML model’s original features does not leak any information to the target VAL domain. Whether sharing the details of the feature engineering is allowed or not, depends on the business agreement among VAL service providers.

[0072] Regardless of how the candidate ML model(s) are found, and how those model(s) are built, the first AIML enablement (first ADAE server 412A with AIML services 502) may be configured to send a (second) ML model retrieval response to the second AIML enablement server (second ADAE server 412B with AIML services) that includes the candidate ML model(s). The second AIML enablement server may be configured to receive the ML model retrieval response, and perform a TL operation in which a source ML model selected out of the candidate ML model(s) is tuned using training data from the target VAL domain 402B to build a target ML model. In some examples, the source ML model includes metadata regarding the source ML model, and the metadata includes e.g., an indication of the number of samples of training data used to train the source ML model. In some of these examples, the source ML model may be tuned based on the number of samples of the training data in the source VAL domain, and a corresponding number of samples of the training data from the target VALdomain. Additionally or alternatively, the source ML model may be tuned based on similarity of one or more ML tasks in both the source VAL domain and the target VAL domain.

[0073] To further illustrate some example implementations of the present disclosure, FIGS. 6Aand 6B illustrate a signaling chart 600 of transfer learning between two vertical domains in an ADAE-based deployment, according to some example implementations. As shown in FIG. 6A, the first AIML enablement server (first ADAE server 412A with AFML services 502) or an AIML producer supporting the source VAL domain 402A has at step 601 trained an ML model based on sufficient training data to capture one or more local dynamics of interest. The ML model may include metadata, such as training time, type of task, type of model, number of samples used for the training or the like. The ML model may also include one or more hyperparameters that are model specific, such as number of layers, number of neurons per layer, activation functions, weights, biases or the like. Additionally or alternatively, the ML model may include performance metrics that are task specific, such as root mean square error (RMSE), accuracy, precision, recall, etc., for the training, validation, and / or test sets.

[0074] The first AIML enablement server (first ADAE server 412A with AIML services 502) or other AIML producer in the source VAL domain 402A decides to persist the ML model (e.g., for future inference, fine tuning, sharing with other VAL domains of the same VAL service provider). The first AIML enablement server / AIML producer at step 602 sends an ML model storage / update request to the A-ADRF 416 or other repository function supporting the source VAL domain (this A-ADRF shown as A-ADRF #1) including the information about model hyperparameters, performance metrics, and metadata of the ML model. The same request message may be used to update an existing model in the A-ADRF by the first AIML enablement server / AIML producer.

[0075] On receiving the request, the A-ADRF 416 supporting the source VAL domain 402A performs an authorization check. If authorization is successful, the A-ADRF stores the ML model in a repository; and at step 603, the A-ADRF sends an ML model storage / update response to the first AIML enablement server (first ADAE server 412A with AIML services 502) / AIML producer in the source VAL domain. The ML model storage / update response may include information about the result of the operation. The same response may also beused to acknowledge update of an existing model in the A-ADRF by the AIML enablement server / AIML producer.

[0076] In some circumstances, an ML model may be needed in the target VAL domain 402B, such as due to changes in environmental conditions of the target VAL domain. In particular, for example, an ML model may be needed due to deployment of a new NPN in a factory, a change in speed limit in a geographic region, or deployment of an loT network in a different environment. One or more AIML producers in the target VAL domain 402B start collecting local information at step 604 for training / fine tuning a ML model. These AIML producer(s) may include, for example, a second VAL server 408B (with AIML services 502) in the second VAL domain, or a second AIML enablement client (second ADAE client 410B with AIML services) or second AIML enablement server (second ADAE server 412B with AIML services) supporting the second VAL domain.

[0077] If prompt collection of sufficient data for model training is not feasible, the AIML producer for the second VAL domain 402B (e.g., second VAL server 408B (with AIML services 502), second AIML enablement client (second ADAE client 410B with AIML services), second AIML enablement server (second ADAE server 412B with AIML services)) may at step 605 decide to retrieve candidate model(s) and apply TL techniques to cope with the lack of data for training a ML model from scratch. This may be the case, for example, if there is a local A-ADRF 616 or other local repository function supporting the target VAL domain 402B (this A-ADRF shown as A-ADRF #2) where baseline ML models may be available, or if there is an agreement between VAL service providers to share “learnings.”

[0078] If a local instance of an A-ADRF 616 is available to the target VAL domain 402B, the AIML producer for the second domain at step 606 sends an ML model retrieval request to the local A-ADRF 616 (e.g., through the second AIML enablement server (second ADAE server 412B with AIML services 502)). The ML model retrieval request includes one or more filtering criteria, such as the type of ML model that is needed, the number of inputs and output(s) that are needed to be handled by the ML model, data types in the input and the output(s), performance threshold conditions, or the like.

[0079] On receiving the request, the local A-ADRF 616 for the target VAL domain 402B performs an authorization check. If authorization is successful, the A-ADRF finds a set of one or more candidate ML models matching the filtering criteria in the request (if any). The localA-ADRF at step 607 sends the set of candidate ML model(s) (including, e.g., hyperparameters, metadata, performance information) back to the AIML producer through the second AIML enablement server (second ADAE server 412B with AIML services 502) in an ML model retrieval response. If no ML model matching the criteria is found, the ML model retrieval response may include indicate a corresponding reject cause, and may also include redirection information to one or more other VAL domains of the same VAL service provider or a different one (conditioned to business agreements).

[0080] As shown at step 608 in FIG. 6B, the second AIML enablement server (second ADAE server 412B with AIML services 502) supporting the target VAL domain 402B forwards an ML model retrieval request to the first AIML enablement server (first ADAE server 412A with AIML services) supporting the source VAL domain 402A, such as over an ADAE-E reference point. The ML model retrieval request includes one or more filtering criteria to find an adequate set of one or more candidate ML models in the A-ADRF 416 supporting the source VAL domain 402A.

[0081] On receiving the request, the first AIML enablement server (first ADAE server 412A with AIML services 502) supporting the source VAL domain 402A performs an authorization check. If authorization is successful, the first AIML enablement server at step609 forwards the ML model retrieval request to the A-ADRF 416 including the one or more filtering criteria.

[0082] The A-ADRF 416 supporting the source VAL domain 402A finds a set of candidate ML model(s) matching the filter criteria included the request (if any). The A-ADRF at step610 sends set of candidate ML model(s) (including, e.g., hyperparameters, metadata, performance information) back to the first AIML enablement server (first ADAE server 412A with AIML services 502) in a ML model retrieval response. If no ML model matching the filter criteria is found, the ML model retrieval response may include indicate a corresponding reject cause.

[0083] The first AIML enablement server (first ADAE server 412A with AIML services 502) at step 611 forwards the ML model retrieval response to the second AIML enablement server (second ADAE server 412B with AIML services) including the set of candidate ML model(s).

[0084] If more than one candidate ML model is found during the retrieval process, the AIML producer for the second VAL domain 402B (e.g., second VAL server 408B (with AIML services 502), second AIML enablement client (second ADAE client 410B with AIML services), second AIML enablement server (second ADAE server 412B with AIML services)) at step 612 selects a source ML model using some selecting policy. For example, the AIML producer may select the source ML model based on performance information of the candidate ML model(s) (during training, validation, testing), and fine-tune the selected, source ML model to build a target ML model. The AIML producer fine-tunes the source ML model based on the amount of data collected so far under the new environment conditions. The source ML model may also be fine-tuned based on similarity of the tasks. In this regard, if the source ML model is built for a regression and the local task in the target VAL domain is also a regression task, a greater number of components in the source ML model may be reused.

[0085] If no ML model was found in any of the repositories, the AIML producer for the target VAL domain 402B may continue collecting data for training (triggered in step 604) until a sufficient amount of data (number of samples of data) is collected.

[0086] The AIML producer for the target VAL domain 402B may decide to store or update the fine-tuned, target ML model through the second AIML enablement server (second ADAE server 412B with AIML services 502) into the local A-ADRF 616. The AIML producer may therefore at step 613 send an ML model storage / update request (e.g., through the second AIML enablement server. The ML model storage / update request may include the new metadata of the target ML model, the new hyperparameters, and the new performance metrics on the training / validation / test sets.

[0087] After an authorization check, the A-ADRF 616 stores or updates the target ML model in the local repository; and the A-ADRF at step 614 sends an ML model storage / update response to the AIML producer (e.g., through the second AIML enablement server (second ADAE server 412B with AIML services 502).

[0088] FIGS. 7A and 7B are flowcharts illustrating various steps in a method 700 according to various example implementations. The method includes sending a machine learning (ML) model retrieval request to a server supporting a source vertical application layer (VAL) domain, as shown at block 702 of FIG. 7A. The method includes receiving a ML model retrieval response from the server based on the ML model retrieval request, the ML modelretrieval response including at least one candidate ML model built using training data from the source VAL domain, as shown at block 704. And the method includes performing a transfer learning operation in which a source ML model selected from the at least one candidate ML model is tuned using training data from a target VAL domain to build a target ML model, as shown at block 706.

[0089] In some examples, the ML model retrieval request is sent at block 702 to the server supporting the source VAL domain that is included in a trust domain of the target VAL domain or to a VAL domain for which a business agreement for sharing models is in place.

[0090] In some examples, the ML model retrieval request includes one or more filtering criteria, and the at least one candidate ML model included in the ML model retrieval response is found by the server based on the one or more filtering criteria.

[0091] In some examples, the at least one candidate ML model included in the ML model retrieval response is built using training data from the source VAL domain that is at least one of normalized, anonymized or obfuscated to protect sensitive information of the source VAL domain from the target VAL domain.

[0092] In some examples, the source ML model includes metadata regarding the source ML model, and the metadata includes an indication of a number of samples of training data from the source VAL domain used to train the source ML model. In some of these examples, the source ML model is tuned based on at least one of the number of samples of the training data from both the source VAL domain and the target VAL domain, or similarity of one or more ML tasks in both the source VAL domain and the target VAL domain.

[0093] In some examples, the ML model retrieval request includes one or more filtering criteria. In some of these examples, the method 700 further includes making a determination that a repository function supporting the target VAL domain is unable to find at least one matching ML model based on the one or more filtering criteria, as shown at block 708 of FIG. 7B. In some of these examples, the ML model retrieval request is sent at block 702 based on the determination.

[0094] FIGS. 8 A and 8B are flowcharts illustrating various steps in a method 800 according to various example implementations. The method includes receiving a machine learning (ML) model retrieval request from a server supporting a target vertical application layer (VAL) domain, as shown at block 802 of FIG. 8A. The method includes retrieving atleast one candidate ML model based on the ML model retrieval request, the at least one candidate ML model built using training data from a source VAL domain, as shown at block 804. And the method includes sending a ML model retrieval response to the server that includes the at least one candidate ML model, the ML model retrieval response sent to enable a transfer learning operation in which a source ML model selected from the at least one candidate ML model is tuned using training data from the target VAL domain to build a target ML model, as shown at block 806.

[0095] In some examples, the ML model retrieval request is received at block 802 from the server supporting the target VAL domain that is included in a trust domain of the source VAL domain.

[0096] In some examples, the ML model retrieval response is a second ML model retrieval response, and retrieving the at least one candidate ML model at block 804 includes sending a second ML model retrieval request to a repository function supporting the source VAL domain, as shown at block 808 of FIG. 8B. And retrieving the at least one candidate ML model includes receiving a first ML model retrieval response from the repository function based on the second ML model retrieval request, the first ML model retrieval response including at least one candidate ML model, as shown at block 810.

[0097] In some examples, the ML model retrieval request includes one or more filtering criteria, and the at least one candidate ML model is retrieved at block 804 based on the one or more filtering criteria.

[0098] In some examples, the at least one candidate ML model that is retrieved at block 804 is built using training data from the source VAL domain that is at least one of normalized, anonymized or obfuscated to protect sensitive information of the source VAL domain from the target VAL domain.

[0099] In some examples, the source ML model includes metadata regarding the source ML model, and the metadata includes an indication of a number of samples of training data from the source VAL domain used to train the source ML model to enable tuning of the source ML model based on at least one of the number of samples of the training data from both the source VAL domain and the target VAL domain, or similarity of one or more ML tasks in both the source VAL domain and the target VAL domain.

[0100] According to example implementations of the present disclosure, a telecommunications system 100 or PLMN 102, and its components such as a UE 110, CN 106, RAN 108, 5GC 202, NG-RAN 204, gNB 206, UE 208, AMF 302, SMF 304, UPF 306, NEF 308, NWDAF 310, AF 312, AS 314, VAL client 404, VAL UE 406, VAL server 408, ADAE client 410, ADAE server 412, A-DCCF 414 and / or A-ADRF 416 may be implemented by various means. Means for implementing the system and its components may include hardware, firmware, software, or combinations thereof. In some examples, one or more apparatuses may be configured to function as or otherwise implement the system and its components shown and described herein. In examples involving more than one apparatus, the respective apparatuses may be connected to or otherwise in communication with one another in a number of different manners, such as directly or indirectly via a wired or wireless network or the like.

[0101] According to some example implementations, at least some of the method 700 described with respect to FIGS. 7A and 7B may be carried out by an apparatus comprising means for performing functions corresponding steps of the method. Similarly, at least some of the method 800 described with respect to FIGS. 8 A and 8B may be carried out by an apparatus comprising means for performing functions corresponding steps of the method. Examples of a suitable apparatus may include an NF (e.g., AMF, SMF, UPF, NEF, NWDAF, AF, AS), VAL client, VAL UE, VAL server, ADAE client, ADAE server, A-DCCF, A-ADRF or any suitable apparatus, such as a server, host or node.

[0102] FIG. 9 illustrates an apparatus 900 in which means for performing various functions includes hardware, alone or under direction of one or more computer programs from a computer-readable storage medium or other memory, such as computer memory, according to some example implementations of the present disclosure. The apparatus may include one or more of each of a number of components such as, for example, processing circuitry 902 connected to computer-readable storage medium or other memory 904.

[0103] The processing circuitry 902 may be composed of one or more processors alone or in combination with one or more computer-readable storage media. The processing circuitry is generally any piece of computer hardware that is capable of processing information such as, for example, data, computer programs and / or other suitable electronic information. The processing circuitry is composed of a collection of electronic circuits some of which may be packaged as an integrated circuit or multiple interconnected integrated circuits (an integratedcircuit at times more commonly referred to as a “chip”). The processing circuitry may be configured to execute computer programs, which may be stored onboard the processing circuitry or otherwise stored in the memory 904 (of the same or another apparatus).

[0104] The processing circuitry 902 may be a number of processors, a multi-core processor or some other type of processor, depending on the particular implementation. Further, the processing circuitry may be implemented using a number of heterogeneous processor systems in which a main processor is present with one or more secondary processors on a single chip. As another illustrative example, the processing circuitry may be a symmetric multi-processor system containing multiple processors of the same type. In yet another example, the processing circuitry may be embodied as or otherwise include one or more ASICs, FPGAs or the like. Thus, although the processing circuitry may be capable of executing a computer program to perform one or more functions, the processing circuitry of various examples may be capable of performing one or more functions without the aid of a computer program. In either instance, the processing circuitry may be appropriately programmed to perform functions or operations according to example implementations of the present disclosure.

[0105] The memory 904 is generally any piece of computer hardware that is capable of storing information such as, for example, data, computer programs, instructions 906 (e.g., computer-readable program code) and / or other suitable information either on a temporary basis and / or a permanent basis. The memory may include volatile and / or non-volatile memory, and may be fixed or removable. Examples of suitable memory include recording media, random access memory (RAM), read-only memory (ROM), a hard drive, a flash memory, a thumb drive, a removable computer diskette, an optical disk or some combination thereof.

[0106] The memory 904 is a non-transitory device capable of storing information. One example of a suitable memory is a computer-readable storage medium, which is distinguishable from a computer-readable transmission medium capable of carrying information from one location to another. Examples of suitable computer-readable transmission media comprise electronic carrier signals, telecommunications signals, software distribution packages, or some combination thereof. As used herein, the term “non-transitory” is a limitation of the medium itself (i.e., tangible, not a signal) as opposed to a limitation on data storage persistency (e.g., RAM versus ROM). A computer-readable medium as describedherein generally refers to a computer-readable storage medium or computer-readable transmission medium. A computer-readable medium is any entity or device capable in which information, such as one or more computer programs or portions thereof, may be stored and carried.

[0107] In addition to the memory 904 (e.g., computer-readable storage medium), the processing circuitry 902 may also be connected to one or more interfaces for displaying, transmitting and / or receiving information. The interfaces may include a communications interface 908 and / or one or more user interfaces (e.g., display, user input interface). The communications interface may be configured to transmit and / or receive information, such as to and / or from other apparatus(es), network(s) or the like. The communications interface may be configured to transmit and / or receive information by physical (wired) and / or wireless communications links. Examples of suitable communication interfaces include a network interface controller (NIC), wireless NIC (WNIC) or the like.

[0108] Execution of the instructions 906 by the processing circuitry 902, or storage of the instructions in the memory 904, supports combinations of operations for implementing example implementations of the present disclosure. In this manner, an apparatus 900 may comprise at least one processing circuitry and at least one memory coupled to the at least one processing circuitry, where the at least one processing circuitry is configured to execute instructions stored in the at least one memory. It will also be understood that one or more functions, and combinations of functions, may be implemented by special purpose hardwarebased computer systems and / or processing circuitry which perform the specified functions, or combinations of special purpose hardware and program code instructions.

[0109] Some example implementations of the present disclosure may also be carried out in the form of a computer process defined by one or more computer programs or portions thereof. Example implementations of the present disclosure may be carried out by executing at least one portion of a computer program comprising instructions. The computer program may be in source code form, object code form, or in some intermediate form. The computer program may be stored in a computer-readable medium that is readable by a computer, processing circuitry or other suitable apparatus. As indicated above, for example, the computer program may be stored in a memory, such as a computer-readable storage medium.Additionally or alternatively, for example, the computer program may be stored in acomputer-readable transmission medium. The coding of software for carrying out example implementations of the present disclosure is well within the scope of a person of ordinary skill in the art.

[0110] As will be appreciated, any suitable instructions may be loaded onto a computer, a processing circuitry or other programmable apparatus from a memory or a computer-readable medium (e.g., computer-readable storage medium, computer-readable transmission medium) to produce a particular machine, such that the particular machine becomes a means for implementing the functions specified herein. The instructions may also be stored in a computer-readable medium that can direct a computer, a processing circuitry or other programmable apparatus to function in a particular manner to thereby generate a particular machine or particular article of manufacture. In some examples, the instructions stored in the computer-readable medium may produce an article of manufacture, where the article of manufacture becomes a means for implementing functions described herein. The instructions may be retrieved from a computer-readable medium and loaded into a computer, processing circuitry or other programmable apparatus to configure the computer, processing circuitry or other programmable apparatus to execute operations to be performed on or by the computer, processing circuitry or other programmable apparatus.

[0111] Retrieval, loading and execution of instructions comprising program code instructions may be performed sequentially such that one instruction is retrieved, loaded and executed at a time. In some example implementations, retrieval, loading and / or execution may be performed in parallel such that multiple instructions are retrieved, loaded, and / or executed together. Execution of the program code instructions may produce a computer-implemented process such that the instructions executed by the computer, processing circuitry or other programmable apparatus provide operations for implementing functions described herein.

[0112] As explained above and reiterated below, the present disclosure includes, without limitation, the following example implementations.

[0113] Clause 1. An apparatus comprising: at least one memory configured to store instructions; and at least one processing circuitry configured to access the at least one memory, and execute the instructions to cause the apparatus to at least: send a machine learning (ML) model retrieval request to a server supporting a source vertical application layer (VAL) domain; receive a ML model retrieval response from the server based on the MLmodel retrieval request, the ML model retrieval response including at least one candidate ML model built using training data from the source VAL domain; and perform a transfer learning operation in which a source ML model selected from the at least one candidate ML model is tuned using training data from a target VAL domain to build a target ML model.

[0114] Clause 2. The apparatus of clause 1 , wherein the ML model retrieval request is sent to the server supporting the source VAL domain that is included in a trust domain of the target VAL domain.

[0115] Clause 3. The apparatus of clause 1 or clause 2, wherein the ML model retrieval request includes one or more filtering criteria, and the at least one candidate ML model included in the ML model retrieval response is found by the server based on the one or more filtering criteria.

[0116] Clause 4. The apparatus of any of clauses 1 to 3, wherein the at least one candidate ML model included in the ML model retrieval response is built using training data from the source VAL domain that is at least one of normalized, anonymized or obfuscated to protect sensitive information of the source VAL domain from the target V AL domain.

[0117] Clause 5. The apparatus of any of clauses 1 to 4, wherein the source ML model includes metadata regarding the source ML model, and the metadata includes an indication of a number of samples of training data from the source VAL domain used to train the source ML model, and wherein the source ML model is tuned based on at least one of the number of samples of the training data from both the source VAL domain and the target VAL domain, or similarity of one or more ML tasks in both the source VAL domain and the target VAL domain.

[0118] Clause 6. The apparatus of any of clauses 1 to 5, wherein the ML model retrieval request includes one or more filtering criteria, and the at least one processing circuitry is configured to execute the instructions to cause the apparatus to further make a determination that a repository function supporting the target VAL domain is unable to find at least one matching ML model based on the one or more filtering criteria, and wherein the ML model retrieval request is sent based on the determination.

[0119] Clause 7. An apparatus comprising: means for sending a machine learning (ML) model retrieval request to a server supporting a source vertical application layer (VAL) domain; means for receiving a ML model retrieval response from the server based on the MLmodel retrieval request, the ML model retrieval response including at least one candidate ML model built using training data from the source VAL domain; and means for performing a transfer learning operation in which a source ML model selected from the at least one candidate ML model is tuned using training data from a target VAL domain to build a target ML model.

[0120] Clause 8. The apparatus of clause 7, wherein the ML model retrieval request is sent to the server supporting the source VAL domain that is included in a trust domain of the target VAL domain.

[0121] Clause 9. The apparatus of clause 7 or clause 8, wherein the ML model retrieval request includes one or more filtering criteria, and the at least one candidate ML model included in the ML model retrieval response is found by the server based on the one or more filtering criteria.

[0122] Clause 10. The apparatus of any of clauses 7 to 9, wherein the at least one candidate ML model included in the ML model retrieval response is built using training data from the source VAL domain that is at least one of normalized, anonymized or obfuscated to protect sensitive information of the source VAL domain from the target VAL domain.

[0123] Clause 11. The apparatus of any of clauses 7 to 10, wherein the source ML model includes metadata regarding the source ML model, and the metadata includes an indication of a number of samples of training data from the source VAL domain used to train the source ML model, and wherein the source ML model is tuned based on at least one of the number of samples of the training data from both the source VAL domain and the target VAL domain, or similarity of one or more ML tasks in both the source VAL domain and the target VAL domain.

[0124] Clause 12. The apparatus of any of clauses 7 to 11, wherein the ML model retrieval request includes one or more filtering criteria, and the apparatus further comprises means for making a determination that a repository function supporting the target VAL domain is unable to find at least one matching ML model based on the one or more filtering criteria, and wherein the ML model retrieval request is sent based on the determination.

[0125] Clause 13. A method comprising: sending a machine learning (ML) model retrieval request to a server supporting a source vertical application layer (VAL) domain; receiving a ML model retrieval response from the server based on the ML model retrieval request, the MLmodel retrieval response including at least one candidate ML model built using training data from the source VAL domain; and performing a transfer learning operation in which a source ML model selected from the at least one candidate ML model is tuned using training data from a target VAL domain to build a target ML model.

[0126] Clause 14. The method of clause 13, wherein the ML model retrieval request is sent to the server supporting the source VAL domain that is included in a trust domain of the target VAL domain.

[0127] Clause 15. The method of clause 13 or clause 14, wherein the ML model retrieval request includes one or more filtering criteria, and the at least one candidate ML model included in the ML model retrieval response is found by the server based on the one or more filtering criteria.

[0128] Clause 16. The method of any of clauses 13 to 15, wherein the at least one candidate ML model included in the ML model retrieval response is built using training data from the source VAL domain that is at least one of normalized, anonymized or obfuscated to protect sensitive information of the source VAL domain from the target VAL domain.

[0129] Clause 17. The method of any of clauses 13 to 16, wherein the source ML model includes metadata regarding the source ML model, and the metadata includes an indication of a number of samples of training data from the source VAL domain used to train the source ML model, and wherein the source ML model is tuned based on at least one of the number of samples of the training data from both the source VAL domain and the target VAL domain, or similarity of one or more ML tasks in both the source VAL domain and the target VAL domain.

[0130] Clause 18. The method of any of clauses 13 to 17, wherein the ML model retrieval request includes one or more filtering criteria, and the method further comprises making a determination that a repository function supporting the target VAL domain is unable to find at least one matching ML model based on the one or more filtering criteria, and wherein the ML model retrieval request is sent based on the determination.

[0131] Clause 19. A computer-readable storage medium that is non-transitory and has instructions stored therein that, in response to execution by at least one processing circuitry, causes an apparatus to at least: send a machine learning (ML) model retrieval request to a server supporting a source vertical application layer (VAL) domain; receive a ML modelretrieval response from the server based on the ML model retrieval request, the ML model retrieval response including at least one candidate ML model built using training data from the source VAL domain; and perform a transfer learning operation in which a source ML model selected from the at least one candidate ML model is tuned using training data from a target VAL domain to build a target ML model.

[0132] Clause 20. The computer-readable storage medium of clause 19, wherein the ML model retrieval request is sent to the server supporting the source VAL domain that is included in a trust domain of the target VAL domain.

[0133] Clause 21. The computer-readable storage medium of clause 19 or clause 20, wherein the ML model retrieval request includes one or more filtering criteria, and the at least one candidate ML model included in the ML model retrieval response is found by the server based on the one or more filtering criteria.

[0134] Clause 22. The computer-readable storage medium of any of clauses 19 to 21, wherein the at least one candidate ML model included in the ML model retrieval response is built using training from the source VAL domain data that is at least one of normalized, anonymized or obfuscated to protect sensitive information of the source VAL domain from the target VAL domain.

[0135] Clause 23. The computer-readable storage medium of any of clauses 19 to 22, wherein the source ML model includes metadata regarding the source ML model, and the metadata includes an indication of a number of samples of training data from the source VAL domain used to train the source ML model, and wherein the source ML model is tuned based on at least one of the number of samples of the training data from both the source VAL domain and the target VAL domain, or similarity of one or more ML tasks in both the source VAL domain and the target VAL domain.

[0136] Clause 24. The computer-readable storage medium of any of clauses 19 to 23, wherein the ML model retrieval request includes one or more filtering criteria, and the computer-readable storage medium has further instructions stored therein that, in response to execution by the at least one processing circuitry, causes the apparatus to further make a determination that a repository function supporting the target VAL domain is unable to find at least one matching ML model based on the one or more filtering criteria, and wherein the ML model retrieval request is sent based on the determination.

[0137] Clause 25. An apparatus comprising means for performing the method of any of clauses 13 to 18.

[0138] Clause 26. A computer-readable medium comprising computer-readable program code that, in response to execution by at least one processing circuitry, causes an apparatus to perform the method of any of clauses 13 to 18.

[0139] Clause 27. A computer-readable storage medium comprising computer-readable program code that, in response to execution by at least one processing circuitry, causes an apparatus to perform the method of any of clauses 13 to 18.

[0140] Clause 28. A computer program comprising computer-readable program code that, in response to execution by at least one processing circuitry, causes an apparatus to perform the method of any of clauses 13 to 18.

[0141] Clause 29. An apparatus comprising: at least one memory configured to store instructions; and at least one processing circuitry configured to access the at least one memory, and execute the instructions to cause the apparatus to at least: receive a machine learning (ML) model retrieval request from a server supporting a target vertical application layer (VAL) domain; retrieve at least one candidate ML model based on the ML model retrieval request, the at least one candidate ML model built using training data from a source VAL domain; and send a ML model retrieval response to the server that includes the at least one candidate ML model, the ML model retrieval response sent to enable a transfer learning operation in which a source ML model selected from the at least one candidate ML model is tuned using training data from the target VAL domain to build a target ML model.

[0142] Clause 30. The apparatus of clause 29, wherein the ML model retrieval request is received from the server supporting the target VAL domain that is included in a trust domain of the source VAL domain.

[0143] Clause 31. The apparatus of clause 29 or clause 30, wherein the ML model retrieval response is a second ML model retrieval response, and the apparatus caused to retrieve the at least one candidate ML model includes the apparatus caused to: send a second ML model retrieval request to a repository function supporting the source VAL domain; and receive a first ML model retrieval response from the repository function based on the second ML model retrieval request, the first ML model retrieval response including at least one candidate ML model.

[0144] Clause 32. The apparatus of any of clauses 29 to 31, wherein the ML model retrieval request includes one or more filtering criteria, and the at least one candidate ML model is retrieved based on the one or more filtering criteria.

[0145] Clause 33. The apparatus of any of clauses 29 to 32, wherein the at least one candidate ML model that is retrieved is built using training data from the source VAL domain that is at least one of normalized, anonymized or obfuscated to protect sensitive information of the source VAL domain from the target VAL domain.

[0146] Clause 34. The apparatus of any of clauses 29 to 33, wherein the source ML model includes metadata regarding the source ML model, and the metadata includes an indication of a number of samples of training data from the source VAL domain used to train the source ML model to enable tuning of the source ML model based on at least one of the number of samples of the training data from both the source VAL domain and the target VAL domain, or similarity of one or more ML tasks in both the source VAL domain and the target VAL domain.

[0147] Clause 35. An apparatus comprising: means for receiving a machine learning (ML) model retrieval request from a server supporting a target vertical application layer (VAL) domain; means for retrieving at least one candidate ML model based on the ML model retrieval request, the at least one candidate ML model built using training data from a source VAL domain; and means for sending a ML model retrieval response to the server that includes the at least one candidate ML model, the ML model retrieval response sent to enable a transfer learning operation in which a source ML model selected from the at least one candidate ML model is tuned using training data from the target VAL domain to build a target ML model.

[0148] Clause 36. The apparatus of clause 35, wherein the ML model retrieval request is received from the server supporting the target VAL domain that is included in a trust domain of the source VAL domain.

[0149] Clause 37. The apparatus of clause 35 or clause 36, wherein the ML model retrieval response is a second ML model retrieval response, and the means for retrieving the at least one candidate ML model includes: means for sending a second ML model retrieval request to a repository function supporting the source VAL domain; and means for receiving a first ML model retrieval response from the repository function based on the second ML modelretrieval request, the first ML model retrieval response including at least one candidate ML model.

[0150] Clause 38. The apparatus of any of clauses 35 to 37, wherein the ML model retrieval request includes one or more filtering criteria, and the at least one candidate ML model is retrieved based on the one or more filtering criteria.

[0151] Clause 39. The apparatus of any of clauses 35 to 38, wherein the at least one candidate ML model that is retrieved is built using training data from the source VAL domain that is at least one of normalized, anonymized or obfuscated to protect sensitive information of the source VAL domain from the target VAL domain.

[0152] Clause 40. The apparatus of any of clauses 35 to 39, wherein the source ML model includes metadata regarding the source ML model, and the metadata includes an indication of a number of samples of training data from the source VAL domain used to train the source ML model to enable tuning of the source ML model based on at least one of the number of samples of the training data from both the source VAL domain and the target VAL domain, or similarity of one or more ML tasks in both the source VAL domain and the target VAL domain.

[0153] Clause 41. A method comprising: receiving a machine learning (ML) model retrieval request from a server supporting a target vertical application layer (VAL) domain; retrieving at least one candidate ML model based on the ML model retrieval request, the at least one candidate ML model built using training data from a source VAL domain; and sending a ML model retrieval response to the server that includes the at least one candidate ML model, the ML model retrieval response sent to enable a transfer learning operation in which a source ML model selected from the at least one candidate ML model is tuned using training data from the target VAL domain to build a target ML model.

[0154] Clause 42. The method of clause 41, wherein the ML model retrieval request is received from the server supporting the target VAL domain that is included in a trust domain of the source VAL domain.

[0155] Clause 43. The method of clause 41 or clause 42, wherein the ML model retrieval response is a second ML model retrieval response, and retrieving the at least one candidate ML model includes: sending a second ML model retrieval request to a repository function supporting the source VAL domain; and receiving a first ML model retrieval response fromthe repository function based on the second ML model retrieval request, the first ML model retrieval response including at least one candidate ML model.

[0156] Clause 44. The method of any of clauses 41 to 43, wherein the ML model retrieval request includes one or more filtering criteria, and the at least one candidate ML model is retrieved based on the one or more filtering criteria.

[0157] Clause 45. The method of any of clauses 41 to 44, wherein the at least one candidate ML model that is retrieved is built using training data from the source VAL domain that is at least one of normalized, anonymized or obfuscated to protect sensitive information of the source VAL domain from the target VAL domain.

[0158] Clause 46. The method of any of clauses 41 to 45, wherein the source ML model includes metadata regarding the source ML model, and the metadata includes an indication of a number of samples of training data from the source VAL domain used to train the source ML model to enable tuning of the source ML model based on at least one of the number of samples of the training data from both the source VAL domain and the target VAL domain, or similarity of one or more ML tasks in both the source VAL domain and the target VAL domain.

[0159] Clause 47. A computer-readable storage medium that is non-transitory and has instructions stored therein that, in response to execution by at least one processing circuitry, causes an apparatus to at least: receive a machine learning (ML) model retrieval request from a server supporting a target vertical application layer (VAL) domain; retrieve at least one candidate ML model based on the ML model retrieval request, the at least one candidate ML model built using training data from a source VAL domain; and send a ML model retrieval response to the server that includes the at least one candidate ML model, the ML model retrieval response sent to enable a transfer learning operation in which a source ML model selected from the at least one candidate ML model is tuned using training data from the target VAL domain to build a target ML model.

[0160] Clause 48. The computer-readable storage medium of clause 47, wherein the ML model retrieval request is received from the server supporting the target VAL domain that is included in a trust domain of the source VAL domain.

[0161] Clause 49. The computer-readable storage medium of clause 47 or clause 48, wherein the ML model retrieval response is a second ML model retrieval response, and theapparatus caused to retrieve the at least one candidate ML model includes the apparatus caused to: send a second ML model retrieval request to a repository function supporting the source VAL domain; and receive a first ML model retrieval response from the repository function based on the second ML model retrieval request, the first ML model retrieval response including at least one candidate ML model.

[0162] Clause 50. The computer-readable storage medium of any of clauses 47 to 49, wherein the ML model retrieval request includes one or more filtering criteria, and the at least one candidate ML model is retrieved based on the one or more filtering criteria.

[0163] Clause 51. The computer-readable storage medium of any of clauses 47 to 50, wherein the at least one candidate ML model that is retrieved is built using training data from the source VAL domain that is at least one of normalized, anonymized or obfuscated to protect sensitive information of the source V AL domain from the target VAL domain.

[0164] Clause 52. The computer-readable storage medium of any of clauses 47 to 51, wherein the source ML model includes metadata regarding the source ML model, and the metadata includes an indication of a number of samples of training data from the source VAL domain used to train the source ML model to enable tuning of the source ML model based on at least one of the number of samples of the training data from both the source VAL domain and the target VAL domain, or similarity of one or more ML tasks in both the source VAL domain and the target VAL domain.

[0165] Clause 53. An apparatus comprising means for performing the method of any of clauses 41 to 46.

[0166] Clause 54. A computer-readable medium comprising computer-readable program code that, in response to execution by at least one processing circuitry, causes an apparatus to perform the method of any of clauses 41 to 46.

[0167] Clause 55. A computer-readable storage medium comprising computer-readable program code that, in response to execution by at least one processing circuitry, causes an apparatus to perform the method of any of clauses 41 to 46.

[0168] Clause 56. A computer program comprising computer-readable program code that, in response to execution by at least one processing circuitry, causes an apparatus to perform the method of any of clauses 41 to 46.

[0169] Many modifications and other implementations of the disclosure set forth herein will come to mind to one skilled in the art to which the disclosure pertains having the benefit of the teachings presented in the foregoing description and the associated figures. Therefore, it is to be understood that the disclosure is not to be limited to the specific implementations disclosed and that modifications and other implementations are intended to be included within the scope of the appended claims. Moreover, although the foregoing description and the associated figures describe example implementations in the context of certain example combinations of elements and / or functions, it should be appreciated that different combinations of elements and / or functions may be provided by alternative implementations without departing from the scope of the appended claims. In this regard, for example, different combinations of elements and / or functions than those explicitly described above are also contemplated as may be set forth in some of the appended claims. Although specific terms are employed herein, they are used in a generic and descriptive sense only and not for purposes of limitation.

Claims

WHAT IS CLAIMED IS:

1. An apparatus comprising: means for sending a machine learning (ML) model retrieval request to server supporting a source vertical application layer (VAL) domain; means for receiving a ML model retrieval response from the server based on the ML model retrieval request, the ML model retrieval response including at least one candidate ML model built using training data from the source VAL domain; and means for performing a transfer learning operation in which a source ML model selected from the at least one candidate ML model is tuned using training data from a target VAL domain to build a target ML model.

2. The apparatus of claim 1, wherein the ML model retrieval request is sent to the server supporting the source VAL domain that is included in a trust domain of the target VAL domain.

3. The apparatus of claim 1 or claim 2, wherein the ML model retrieval request includes one or more filtering criteria, and the at least one candidate ML model included in the ML model retrieval response is found by the server based on the one or more filtering criteria.

4. The apparatus of any of claims 1 to 3, wherein the at least one candidate ML model included in the ML model retrieval response is built using training data from the source VAL domain that is at least one of normalized, anonymized or obfuscated to protect sensitive information of the source VAL domain from the target VAL domain.

5. The apparatus of any of claims 1 to 4, wherein the source ML model includes metadata regarding the source ML model, and the metadata includes an indication of a number of samples of training data from the source VAL domain used to train the source ML model, and wherein the source ML model is tuned based on at least one of the number of samples of the training data from both the source VAL domain and the target VAL domain, orsimilarity of one or more ML tasks in both the source VAL domain and the target VAL domain.

6. The apparatus of any of claims 1 to 5, wherein the ML model retrieval request includes one or more filtering criteria, and the apparatus further comprises means for making a determination that a repository function supporting the target VAL domain is unable to find at least one matching ML model based on the one or more filtering criteria, and wherein the ML model retrieval request is sent based on the determination.

7. A method comprising: sending a machine learning (ML) model retrieval request to a server supporting a source vertical application layer (VAL) domain; receiving a ML model retrieval response from the server based on the ML model retrieval request, the ML model retrieval response including at least one candidate ML model built using training data from the source VAL domain; and performing a transfer learning operation in which a source ML model selected from the at least one candidate ML model is tuned using training data from a target VAL domain to build a target ML model.

8. The method of claim 7, wherein the ML model retrieval request is sent to the server supporting the source VAL domain that is included in a trust domain of the target VAL domain.

9. The method of claim 7 or claim 8, wherein the ML model retrieval request includes one or more filtering criteria, and the at least one candidate ML model included in the ML model retrieval response is found by the server based on the one or more filtering criteria.

10. The method of any of claims 7 to 9, wherein the at least one candidate ML model included in the ML model retrieval response is built using training data from the source VAL domain that is at least one of normalized, anonymized or obfuscated to protect sensitive information of the source VAL domain from the target VAL domain.

11. The method of any of claims 7 to 10, wherein the source ML model includes metadata regarding the source ML model, and the metadata includes an indication of a number of samples of training data from the source VAL domain used to train the source ML model, and wherein the source ML model is tuned based on at least one of the number of samples of the training data from both the source VAL domain and the target VAL domain, or similarity of one or more ML tasks in both the source VAL domain and the target VAL domain.

12. The method of any of claims 7 to 11, wherein the ML model retrieval request includes one or more filtering criteria, and the method further comprises making a determination that a repository function supporting the target VAL domain is unable to find at least one matching ML model based on the one or more filtering criteria, and wherein the ML model retrieval request is sent based on the determination.

13. An apparatus comprising: means for receiving a machine learning (ML) model retrieval request from a server supporting a target vertical application layer (VAL) domain; means for retrieving at least one candidate ML model based on the ML model retrieval request, the at least one candidate ML model built using training data from a source VAL domain; and means for sending a ML model retrieval response to the server that includes the at least one candidate ML model, the ML model retrieval response sent to enable a transfer learning operation in which a source ML model selected from the at least one candidate ML model is tuned using training data from the target VAL domain to build a target ML model.

14. The apparatus of claim 13, wherein the ML model retrieval request is received from the server supporting the target VAL domain that is included in a trust domain of the source VAL domain.

15. The apparatus of claim 13 or claim 14, wherein the ML model retrieval response is a second ML model retrieval response, and the means for retrieving the at least one candidate ML model includes: means for sending a second ML model retrieval request to a repository function supporting the source VAL domain; and means for receiving a first ML model retrieval response from the repository function based on the second ML model retrieval request, the first ML model retrieval response including at least one candidate ML model.

16. The apparatus of any of claims 13 to 15, wherein the ML model retrieval request includes one or more filtering criteria, and the at least one candidate ML model is retrieved based on the one or more filtering criteria.

17. The apparatus of any of claims 13 to 16, wherein the at least one candidate ML model that is retrieved is built using training data from the source VAL domain that is at least one of normalized, anonymized or obfuscated to protect sensitive information of the source VAL domain from the target VAL domain.

18. The apparatus of any of claims 13 to 17, wherein the source ML model includes metadata regarding the source ML model, and the metadata includes an indication of a number of samples of training data from the source VAL domain used to train the source ML model to enable tuning of the source ML model based on at least one of the number of samples of the training data from both the source VAL domain and the target VAL domain, or similarity of one or more ML tasks in both the source VAL domain and the target VAL domain.

19. A method comprising: receiving a machine learning (ML) model retrieval request from a server supporting a target vertical application layer (VAL) domain; retrieving at least one candidate ML model based on the ML model retrieval request, the at least one candidate ML model built using training data from a source VAL domain; andsending a ML model retrieval response to the server that includes the at least one candidate ML model, the ML model retrieval response sent to enable a transfer learning operation in which a source ML model selected from the at least one candidate ML model is tuned using training data from the target VAL domain to build a target ML model.

20. The method of claim 19, wherein the ML model retrieval request is received from the server supporting the target VAL domain that is included in a trust domain of the source VAL domain.

21. The method of claim 19 or claim 20, wherein the ML model retrieval response is a second ML model retrieval response, and retrieving the at least one candidate ML model includes: sending a second ML model retrieval request to a repository function supporting the source VAL domain; and receiving a first ML model retrieval response from the repository function based on the second ML model retrieval request, the first ML model retrieval response including at least one candidate ML model.

22. The method of any of claims 19 to 21, wherein the ML model retrieval request includes one or more filtering criteria, and the at least one candidate ML model is retrieved based on the one or more filtering criteria.

23. The method of any of claims 19 to 22, wherein the at least one candidate ML model that is retrieved is built using training data from the source VAL domain that is at least one of normalized, anonymized or obfuscated to protect sensitive information of the source VAL domain from the target VAL domain.

24. The method of any of claims 19 to 23, wherein the source ML model includes metadata regarding the source ML model, and the metadata includes an indication of a number of samples of training data from the source VAL domain used to train the source ML model to enable tuning of the source ML model based on at least one of the number of samples of thetraining data from both the source VAL domain and the target VAL domain, or similarity of one or more ML tasks in both the source VAL domain and the target VAL domain.

Images