R1-2410335
discussion
Other Aspects of AI/ML framework
From AT&T
Summary
AT&T's contribution discusses the general framework for AI/ML in NR air interface, focusing on unified lifecycle management (LCM), model identification procedures, and model transfer mechanisms. The document contains 13 detailed proposals covering framework unification, model identification options, and standardized model transfer approaches.
Position
AT&T advocates FOR a unified LCM framework that combines functionality-based and model-ID-based operations with functionality as default, network-controlled model ID assignment, and standardized model transfer approaches focusing on known model structures. They push AGAINST fragmented LCM approaches and advocate for practical, implementable solutions that avoid proprietary disclosure risks and excessive cross-vendor collaboration burdens.
Key proposals
- Proposal 1 (Sec 2): Support a unified LCM providing both functionality-based and model-ID-based operations, with functionality-based operation supported by default and model-ID used when needed
- Proposal 2 (Sec 2): Confirm definitions for supported functionalities (UE capability via RRC/LPP), applicable functionalities (UE ready for inference), and activated functionalities (enabled for inference)
- Proposal 3 (Sec 3): For all model identification options (MI-option 1-4) for type B, network assigns the model ID(s) for identified models if assignment is needed
- Proposal 4 (Sec 3): Study four relationship options between model ID(s) and associated ID(s) in AI-Example1, including one-to-one, one-to-many, many-to-one, and many-to-many mapping
- Proposal 5 (Sec 3): Study and down-select alternatives for determining/assigning model ID including NW assignment, UE assignment, associated ID as model ID, or pre-defined specification rules
- Proposal 6 (Sec 3): Prioritize study of Option 1 (NW assigns Model ID) and Option 3 (Associated ID assumed as model ID) for Rel-19 assessment
- Proposal 7 (Sec 3): For MI-Option2 with dataset transfer in two-sided models, UE-part cannot be identified by dataset ID-X alone; model ID or pairing ID is required
- Proposal 8 (Sec 3): For MI-Option2 dataset transfer, transmit input data, labels, their associations, format/type information, and dataset size from network to UE
- Proposal 9 (Sec 3): For MI-Option4 with multiple standardized reference models, pre-define associated model ID for each reference model for identification
- Proposal 10 (Sec 4): Support model transfer/delivery for both UE-sided models and UE-part of two-sided models in Rel-19
- Proposal 11 (Sec 4): For model transfer Case z4, RAN1 focuses on standardized known model structures option, as offline alignment is beyond RAN1 expertise
- Proposal 12 (Sec 4): Prioritize study of standardized known model structures for UE-sided model/UE part of two-sided model in Case z4
- Proposal 13 (Sec 4): Study three options for open format in Case z4: reuse existing AI community formats (ONNX), define new 3GPP format (ASN.1), or reuse SA2 interoperability token mechanism