Content for TS 23.482 Word version: 19.3.0
5 Application architecture for enabling AI/ML services
5.1 General
5.2 Application enablement architecture
5.2.1 On-Network AIML Enablement (AIMLE) Functional Architecture
5.2.2 Off-Network AIMLE Functional Architecture
5.2.3 Functional Entities Description
5.2.4 Reference Points Description
6 AIMLE Functional Description
7 Identities and commonly used values
...
...
5 Application architecture for enabling AI/ML services p. 15
5.1 General p. 15
The functional architecture enhancements for the AIML enablement service are based on the generic functional model specified in clause 6.2 of TS 23.434. The architecture enhancements are organized into functional entities to describe a functional architecture enhancement which addresses the support for AIML enablement aspects for vertical applications.
5.2 Application enablement architecture p. 16
5.2.1 On-Network AIML Enablement (AIMLE) Functional Architecture p. 16
Figure 5.2.1-1 illustrates the on-network functional model of AIMLE. In the vertical application layer, the VAL client communicates with the VAL server over VAL-UU reference point. VAL-UU supports both unicast and multicast delivery modes. The AIMLE functional entities on the UE and the server are grouped into AIMLE client(s) and AIMLE server(s) respectively.
The AIMLE includes of a common set of services for comprehensive enablement of AIML functionality, including federated and distributed learning (e.g., FL client registration management, FL client discovery and selection), and reference points. The AIMLE services are offered to the vertical application layer (VAL).
The AIMLE client communicates with the AIMLE server(s) over the AIML-UU reference points. The AIMLE client provides functionality to the VAL client(s) over AIML-C reference point. The VAL server(s) communicate with the AIMLE server(s) over AIML-S reference points. The AIMLE servers communicate with the underlying 3GPP network systems using the respective 3GPP interfaces specified by the 3GPP network system. AIML-E reference point enables interactions between two AIMLE servers (e.g. central and edge AIMLE servers).
The AIMLE server interacts with the ML repository which serves as repository for ML model and ML participants over AIML-R.
5.2.1.1 Service-based AIMLE architecture representation p. 16
Figure 5.2.1.1-1 exhibits the service-based interfaces for providing and consuming AIMLE services. The AIMLE server could provide service to VAL server and AIMLE client through interface SAiml.
Figure 5.2.1.1-1: Architecture for AIML enablement - Service based representation.
(⇒ copy of original 3GPP image)
(⇒ copy of original 3GPP image)
Figure 5.2.1.1-2 illustrates the service-based representation for utilization of the 5GS network services based on the 5GS SBA specified in TS 23.501.
Figure 5.2.1.1-2: Architecture for AIMLE utilizing the 5GS network services based on the 5GS SBA - Service based representation,
(⇒ copy of original 3GPP image)
(⇒ copy of original 3GPP image)
The AIMLE server as well as ADAES is deployed as a SEAL server; hence enhancements to SEAL architecture (as specified in TS 23.434) are needed to incorporate the AIMLE service. Figure 5.2.1.1-3 illustrates the service-based representation including AIMLE server as part of the SEAL framework.
Figure 5.2.1.1-3: SEAL functional model representation using service-based interfaces and including AIMLE function.
(⇒ copy of original 3GPP image)
(⇒ copy of original 3GPP image)
5.2.2 Off-Network AIMLE Functional Architecture p. 18
Figure 5.2.2-1 illustrates the off-network (UE-to-UE) functional model of AIML enablement. In the vertical application layer, the VAL client communicates with a further VAL client over VAL-PC5 reference point. VAL-PC5 supports both unicast and multicast delivery modes. The UE1, if connected to the network via Uu reference point, can also act as a UE-to-network relay, to enable UE2 to access the VAL server(s) over the VAL-UU reference point.
The AIMLE client communicates with a further AIMLE client(s) over the AIML-PC5 reference points. The AIMLE client provides functionality to the VAL client(s) over AIML-C reference point. Such communication is performed for supporting local ML operations (training, distribution, inference) in a coordinated manner.
The off network functional architecture is similar to SEAL off-network architecture (as specified in TS 23.434).
5.2.3 Functional Entities Description p. 18
5.2.3.1 General p. 18
Each subclause is a description of a functional entity and does not imply a physical entity.
5.2.3.2 AIMLE client p. 18
The AIMLE client functional entity acts as the application client supporting AIMLE services. It interacts with the AIMLE server.
5.2.3.3 AIMLE server p. 18
The AIMLE server functional entity provides AIMLE services supported within the vertical application layer. It interacts with the AIMLE client, the 3GPP network, other SEAL services, and VAL server.
5.2.3.4 ML repository p. 18
The ML repository is a logical entity that serves both as a registry for ML/FL members and as a repository for application layer ML model related information. It can be accessed by the AIMLE server.
5.2.4 Reference Points Description p. 19
5.2.4.1 General p. 19
The reference points for the functional model related to AIMLE are described in the following subclauses.
5.2.4.2 AIML-UU p. 19
The interactions related to AIML enablement functions between the AIMLE client and AIMLE server are supported by AIML-UU reference point. This reference point utilizes Uu reference point as described in TS 23.401 and TS 23.501.
5.2.4.3 AIML-S p. 19
The interactions related to AIML enablement functions between the VAL server(s) and the AIMLE server are supported by AIML-S reference point.
5.2.4.4 AIML-C p. 19
The interactions related to AIML enablement functions between the VAL client(s) and the AIMLE client within a VAL UE are supported by AIML-C reference point.
5.2.4.5 AIML-R p. 19
The interactions related to AIML enablement functions between the AIMLE server, and the ML repository are supported by AIML-R reference point.
5.2.4.6 AIML-E p. 19
The interactions related to AIML enablement functions between AIMLE servers (e.g., central and edge AIMLE servers).
5.2.4.7 AIML-PC5 p. 19
The interactions related to AIML enablement functions between AIMLE clients in off-network deployments.
6 AIMLE Functional Description p. 19
6.1 Support for ML model retrieval p. 19
This functionality covers the retrieval of ML models by an AIMLE client or a VAL server via the AIMLE server. The AIMLE server provides the functionality required to retrieve ML models stored in a ML repository based on a filtering criteria. The functionality is provided to the ML model consumer via an API following a request/response model or a subscribe/notify model.
6.2 Support for ML model training p. 19
This functionality enables the AIMLE server to support ML model training based on requests from a VAL server. The AIMLE server provides the functionality to train a specific ML model or assist the ML model training at VAL server/client(s).
6.3 Support for FL member registration p. 20
This functionality covers the registration, registration update and de-registration of the candidate FL member to the ML repository which is keeping the FL member registrations. Such candidate member can be a VAL server functionality or an enabler layer functionality (e.g. AIMLE server) which is registering to the ML repository/registry to act as FL member for a given application event (analytics event or event triggered by a VAL layer application server).
6.4 Support for FL events subscription and notification p. 20
This functionality enables a consumer (who can be the AIMLE server or a VAL server e.g. acting as FL server) to subscribe for FL related events and getting notified on changes on the availability of the FL members which are to be used for the FL-related task (e.g., training). This capability at the ML repository acting as an AIML service registry supports the subscription for events related to FL members and the notification to the consumer in case of changes. This feature assumes that such FL members (AIMLE or VAL server or AIMLE clients) have previously registered to this registry their availability and capabilities.
6.5 Support AI/ML task transfer p. 20
This functionality covers the AIMLE support for ML task transfer, which is applicable to scenarios where an AI/ML member cannot finish the assigned AI/ML task during the performing process. In this scenario, the AIMLE server assists the source AI/ML member by support transferring the intermediate AI/ML information (e.g., the intermediate AI/ML operation status and results) to another AI/ML member (target AI/ML member) for further operations to complete the AI/ML task.
6.6 Support for AIMLE client registration p. 20
This functionality allows an AIMLE client (e.g., AI/ML capable UEs) to register with an AIMLE server. The AIMLE server stores the client information for future interactions. This functionality is crucial for enabling the AIMLE client to participate in AI/ML operations.
6.7 Support for AIMLE client discovery p. 20
This functionality enables the VAL server to discover available AIMLE clients for AI/ML operations, such as training or inference. The AIMLE server provides the functionality to select suitable AIMLE clients that fulfill the discovery criteria.
6.8 Support for AIMLE client selection p. 20
This functionality enables the selection of AIMLE clients to participate in AI/ML operations. There are two modes for client selection: VAL server selection and AIMLE server selection. In VAL server selection, the functionality is provided by the AIMLE Server selecting candidate AIMLE clients from the client list provided by the VAL Server. In AIMLE server selection, the functionality is provided by the AIMLE Server selecting candidate AIMLE clients based on the client selection criteria provided by the VAL Server from the AIMLE clients in the ML repository.
6.9 Support for AIMLE client participation p. 20
This functionality enables the AIMLE server to verify and manage the participation of AIMLE clients in AI/ML operations. The AIMLE client responds with its willingness to perform AI/ML operations based on the information provided by the AIMLE server.
6.10 Support for ML model management p. 20
This functionality enables the AIMLE server to manage ML models through interaction with the ML repository. The AIMLE server provides the storage and discovery functionality for ML model information. The storage functionality allows the AIMLE server to store ML model information in the repository. The discovery functionality enables the AIMLE server to retrieve information about available ML models via an API following a request/response model.
6.11 Support HFL training p. 21
This functionality provides the AIMLE server support for horizontal federated learning. This support is applicable to the case where multiple AIMLE clients are expected to locally train the ML model, and the AIMLE server is required to select, configure and coordinate the HFL clients.
6.12 Support AIMLE client selection subscription and notification p. 21
This functionality is related to the AIMLE client selection subscription request and notification to enable VAL Servers to subscribe for monitoring AIML members who meet criteria for performing an AIMLE service, selecting AIMLE clients and receiving notification when there is an update on the selected and re-selected AIML client's status when re-selection is performed according to AIML member selection criteria.
6.13 Support for Split AI/ML Operation p. 21
Split AI/ML operation is a type of AI/ML operation that allows distributed processing related to ML models into multiple stages on different processing nodes. The intention is to offload the computation-intensive and energy-intensive AI/ML stages to network endpoints, whereas leave the privacy-sensitive and delay-sensitive stage at the end device as described in TS 22.261.
The following functionality is provided to support split AI/ML operation:
- An application consuming services from the AI/ML application enablement layer can discover or manage (e.g., create, update, delete) a split operation profile with the AIMLE server for the purpose of consuming results from corresponding instance of a split AI/ML operation pipeline.
- A VAL server can register with the AIMLE server to indicate its capabilities for acting as a processing node of an instance of a split AI/ML operation pipeline.
- An application consuming services from the AI/ML application enablement layer can subscribe with the AIMLE server to receive event notifications related to an instance of a split AI/ML operation pipeline.
6.14 Support data management assistance p. 21
This functionality covers the AIMLE assistance in data management related operations in the ML model lifecycle. AIMLE data management assistance is the process of the AIMLE server assisting AIMLE service consumers with managing data operations (data preparation and processing) performed by VAL clients.
6.15 Support for Transfer Learning enablement p. 21
This functionality covers the support for discovering and selecting pre-trained models transfer learning operations in application enablement layer, where the support is based on the request for either an ML task from VAL layer or for an analytics task from ADAES. Transfer Learning enablement allows the consumer to discover the similar ML models to be used as base models for the TL, as well as to support the selection of the best model to be used as pre-trained model.
6.16 Support for FL member grouping p. 21
This functionality covers the AIMLE capability to enable the group management of the entities serving as FL clients at the application enablement layer. Such group management is about the creation, monitoring and update of the FL member groups based on the AI/ML operations, which are based on 1) the analytics event/service by ADAES or 2) the VAL requirement for FL support services.
6.17 Support vertical federated learning p. 22
This functionality is related to the AIMLE support for vertical FL (VFL) among AIMLE clients serving as VFL members. This capability involves the determination of employing VFL based on the ML model training request, the feature alignment and the decision on the AIMLE clients serving as VFL clients, based on the training capability evaluation functionality.
6.18 Support for ML model training capability evaluation p. 22
This functionality enables AIMLE server to request AIMLE client for ML model training capability evaluation to support FL training (e.g. HFL, VFL). AIMLE client evaluates its capability and availability to join the FL training process and responds to the AIMLE server with the evaluation result (e.g. join the FL training process with test result, or not join with fail reason). The ML model training capability evaluation result can be used by the AIMLE server to select FL members for FL training process (e.g. HFL, VFL).
6.19 Support AIML service operations control and management p. 22
This functionality enables the VAL server (as AIMLE service consumer) to control the operation mode of an AIMLE service for a given AI/ML operation.
6.20 Support for ML model update p. 22
This functionality enables the AIMLE server to update trained and deployed ML models by detecting model performance degradation, and triggering ML model re-training and update to fix the observed degradation.
6.21 Support for ML model performance monitoring p. 22
This functionality covers a capability of AIMLE server for monitoring detecting a degradation relation to an ML operation / analytics operation and translating to an ML model degradation (expected or predicted) and performing an action to alleviate this issue (new model training or re-training).
6.22 Support for AIMLE assisted ML model selection p. 22
This functionality enables an AIMLE service consumer to request assistance with ML model selection from an AIMLE server. The AIMLE server returns a list of ML models with corresponding model performance.
6.23 Support for AIMLE context transfer in edge data networks p. 22
This functionality provides support for AIMLE operations performed by AIMLE clients spread across multiple edge service areas in edge data networks. The functionality allows the context transfer between edge AIMLE servers when an AIMLE client moves to a different edge service area.
6.24 Support for Assisting Hierarchical Computing p. 22
This functionality provides support for assisting hierarchical computing by AIMLE servers. The functionality allows the consumer to request hierarchical computing assistance information to support its operations.
7 Identities and commonly used values p. 23
7.1 General p. 23
The common identities for SEAL and ADAES refer to TS 23.434 and TS 23.436 respectively. The following clauses list the additional identities and commonly used values for AIMLE Service.
7.2 AIMLE server ID p. 23
The AIMLE server ID uniquely identifies the AIML enablement server.
7.3 AIMLE client ID p. 23
The AIMLE client ID uniquely identifies the AIML enablement client.
7.4 ML repository ID p. 23
The ML repository ID uniquely identifies the ML-related repository function, which is used for storing the ML models and ML related information, as well as for serving as registry for the ML / FL members.
7.5 ML model ID p. 23
The ML model ID uniquely identifies the application-layer ML model.
7.6 FL member ID p. 23
The FL member ID uniquely identifies the participant entity in a Federated Learning process which is supported by AIMLE service, e.g., AIMLE server ID, VAL server ID or EAS ID.
7.7 AIMLE service area p. 23
The AIMLE service area is the area where the AIML Enablement server owner provides its AIML support services. The AIMLE service area can be expressed as a Topological Service Area (e.g., a list of TA), a Geographical Service Area (e.g., geographical coordinates) or both.
7.8 ML model profile ID p. 23
The ML model profile ID uniquely identifies a ML model profile that is created by the ML repository as a result of a successful ML model storage request by a AIMLE server.
7.9 AIMLE client set ID p. 23
The AIMLE client set ID is a unique identifier assigned by an AIMLE server that is associated with the set of AIMLE clients that have been selected and have agreed to participate in performing AI/ML related operations, e.g., ML model training.
