Content for  TR 22.874  Word version:  18.2.0

The AI/ML models can be classified and generated from different perspectives, such as services, users, time, locations, etc. So there can be a large number of AI/ML models (incl. data for structure and weight factors, gradient) used for inference and training. Due to the limitation of storage, UE cannot always preload all AI/ML models for different works. It can be a new commercial opportunity that operators provide services to help manage and distribute the AI/ML models so that UE can acquire a proper model immediately. In the future, it is expected that 3rd party companies will make use of AI models to support different kinds of services such as panorama tourist guide using Augmented Reality (AR) in a resort. However, most of 3rd party does not have the resource (server) in distributed places. Considering the large size of AI models, strict downloading time (illustrated in use case 6.1, 6.2, 6.3) and limited UE storage, it is appropriate for 3rd party companies to authorize professional companies to manage their AI models rather than doing the same thing locally by themselves. Since operator cloud resource has advantage to manage different AI/ML models centrally (e.g. cloud server) and locally (e.g. MEC server), such services can be exposed to 3rd party users as well. Specifically,

• The operator has its own need to maintain a AI/ML model pool for its own business like network optimization;
• The 3rd party would like to use some common AI/ML models already stored in operator's cloud;
• Due to resource limitation of 3rd party, they would like to lease operator's cloud resource to manage their AI models.

Therefore, as shown in Figure 6.4.1-1, The AI model management includes:

• Let 3rd party invoke capabilities exposed by 5GS to upload/download/update/delete/store/monitor AI models.
• Transmitting AI models to user efficiently per situation (e.g. entering into a certain area)
• As the 5GS may collect communication data, user experience, etc., it may perform a data analytics for user experience, which can generate analyzing results to help operator's cloud train and improve AI models for the 3rd party.

The operator cloud (e.g. Edge server) stores a variety of AI/ML models according to operator's or 3rd party's requirement. The operator cloud (e.g. Edge server) is capable to distribute a model stored in the operator cloud to devices.

1. Company A provides panorama tourist guide using Augmented Reality (AR) technology. They provide guide services in a commercial area and a tourist resort. UE needs to download Model A and B (both are 32 bits VGGnet, 536MByte) respectively;
2. To reach the SLA, company A indicates to the operator that UE requires model A in area A and model B in area B.
3. When UE moves to one place, The local Edge server trigger to establish a QoS acceleration and UE downloads the corresponding Model (A or B) stored in the local Edge server timely so that user can enjoy the continuous AR world without obvious interruption when model changes.

UE uses the Model for panorama tourist guide.

None.

[P.R.6.4-001] [P.R.6.4-002]