R1-2409881
discussion
Further study on AI/ML model and data
From Xiaomi
Summary
Xiaomi proposes technical solutions for AI/ML model delivery/transfer and identification in NR air interfaces, arguing for standardized solutions particularly for network-side trained models. The document contains 13 proposals and 10 observations focused on Case z4 model transfer and various model identification options.
Position
Xiaomi advocates FOR network-side AI model training and standardized model transfer/delivery solutions, particularly Case z4 with specified known model structures. They push AGAINST over-the-air dataset delivery and argue for cell-group unique IDs as a compromise between UE complexity and network proprietary deployment preservation. They oppose making model parameter delivery mandatory and advocate for UE consent-based approaches.
Key proposals
- Proposal 1 (Sec 1): Consider standardised solutions for model transfer/delivery at least for the case that AI models are trained on network side
- Proposal 2 (Sec 1): When the AI models are developed by the network side, prioritize investigating model transfer/delivery solution case z4 with specified known model structure
- Proposal 3 (Sec 1): Regarding to model transfer/delivery Case z4, identify the potential need for model parameter delivery and confirm UE's consent on the model parameter delivery
- Proposal 4 (Sec 1): UE should report to network that transferred model parameters are ready for inference
- Proposal 5 (Sec 1): Further study how to define the second indication for Case z4
- Proposal 6 (Sec 2): The associated ID is not equivalent to the model ID
- Proposal 7 (Sec 2): Support cell-group unique associated ID to balance complexity on UE side and proprietary deployment preservation on NW side
- Proposal 8 (Sec 2): How to utilize datasets for UE-part of two-side models development depends on UE implementation
- Proposal 9 (Sec 2): Support cell-group unique ID-X to balance complexity on UE side and proprietary deployment preservation on NW side
- Proposal 10 (Sec 2): Consider 4-step procedure for MI-Option 3 with separate model identification and transfer phases
- Proposal 11 (Sec 2): Support to discuss MI-Option 4 in AI 9.1.4.1 (CSI compression)
- Proposal 12 (Sec 2): Consider detailed 4-step procedure for Type A model identification including offline coordination
- Proposal 13 (Sec 2): Associated ID can be considered for data collection for Type A model identification