R1-2409483
discussion
Discussion on other aspects of AI/ML model and data
From ZTE
ZTE's prior position on
9.1.4.2
at
RAN1#118bis
· AI-synthesized, paraphrased
verify sources →
Strongly opposes MI-Option 2 (dataset transfer) due to resource overhead and feasibility concerns, while strongly supporting MI-Option 3 (model transfer) and MI-Option 4 (standardized reference models).
Summary
ZTE analyzes model identification options for two-sided AI/ML models in NR, arguing against dataset transfer (MI-Option 2) due to high overhead and latency, while favoring model parameter transfer (MI-Option 3) and standardization of reference models (MI-Option 4). The document contains 14 proposals and 8 observations, prioritizing Type B model identification and specific model transfer cases (z4) for Rel-19 studies.
Position
ZTE presents a technical case against MI-Option 2 (dataset transfer), citing huge resource overhead, large latency, and potential performance degradation due to backbone misalignment. They prefer MI-Option 4 (standardization of reference UE-part model) and MI-Option 3 (model transfer), specifically prioritizing Type B model identification and model transfer Case z4 (network-trained parameters for known structure) for two-sided models. ZTE proposes that the dataset ID serves as the model ID in MI-Option 2 and that dataset transfer mechanisms be handled by higher layer signaling under RAN2 scope. They argue that specification impacts for reference model structures should be studied within the CSI compression agenda item to avoid duplication. Finally, they propose studying the timeline for model readiness and the feasibility of partial parameter transfer for Case z4.
Key proposals
- Proposal 1 (Sec 2.1.1): Regarding MI-Option 2, dataset ID is considered as model ID to alleviate LCM burden.
- Proposal 2 (Sec 2.1.1): RAN1 further studies the necessity and potential approaches to deal with the impact of UE-side additional conditions for the dataset.
- Proposal 3 (Sec 2.1.1): Dataset transfer can be realized by higher layer signalling, with detailed mechanisms left to RAN2.
- Proposal 4 (Sec 2.1.3): Regarding MI-Option 4, standardization of reference UE-part model is preferred to resolve multi-vendor collaboration issues.
- Proposal 5 (Sec 2.1.3): Standardization of reference model can be further studied in the CSI compression agenda item to avoid duplicated discussion.
- Proposal 6 (Sec 2.1.4): In Rel-19 AI/ML framework study, type B model identification is prioritized compared with type A model identification.
- Proposal 7 (Sec 2.2): In Rel-19 AI/ML framework study, RAN1 prioritizes the model transfer study for two-sided model rather than UE-side model.
- Proposal 8 (Sec 2.2): In Rel-19 AI/ML framework study, RAN1 prioritizes the model transfer z4 for two-sided model.
- Proposal 9 (Sec 2.2): Regarding model transfer/delivery Case z4, further study the details of the assigned ID associated with the transferred model parameters.
- Proposal 10 (Sec 2.2): The assigned ID associated with transferred parameters can be considered as model ID for subsequent LCM procedure.
- Proposal 11 (Sec 2.2): Further study how to define the timeline as a starting point for when the AI model with transferred parameters is ready for inference.
- Proposal 12 (Sec 2.2): Evaluation study needs to be further performed to verify the feasibility of transferring partial parameters for a known model structure.
- Proposal 13 (Sec 2.2): The details of specification impacts of model structure can be further studied and discussed in the CSI compression agenda item.
- Proposal 14 (Sec 2.3): RAN1’s work on CN/OAM/OTT collection of UE-sided model training data can be triggered by RAN2 LS if needed.