R1-2407657
discussion
Discussion on other aspects of the additional study for AI/ML
From Huawei
Summary
This Huawei document discusses AI/ML air interface aspects focusing on model identification for two-sided models, model transfer/delivery, and UE-side training data collection. The document contains 8 proposals and 3 observations across multiple sections covering different model identification options and deployment scenarios.
Position
Huawei advocates FOR focusing model identification discussions exclusively on two-sided models while eliminating MI-Option 1 from scope, and pushes FOR using Case y as baseline for UE-side training scenarios. They are AGAINST pursuing Case z1 for UE-side trained models due to cross-vendor collaboration burdens and lack of benefits over Case y, and AGAINST prioritizing UE data collection mechanisms at RAN1 since it's outside their scope.
Key proposals
- Proposal 1 (Sec 2): There is no need to further discuss MI-Option 1 since it is applicable only to one-sided model cases, while the scope of model identification in revised WID is limited to two-sided cases
- Proposal 2 (Sec 2.1): For MI-Option 2 dataset transfer from NW to UE, information needed includes content of model input/label for training UE side CSI generation part plus meta information like dataset ID, size, type/format, association between ground-truth CSI and CSI feedback
- Proposal 3 (Sec 2.1): For MI-Option 2 procedure, model identification is achieved when the dataset ID is delivered together with the delivered dataset
- Proposal 4 (Sec 2.2): MI-Option 3 is applicable to two-sided model case
- Proposal 5 (Sec 2.2): For MI-Option 3 transmitted information, Case z4 needs model parameters plus meta information including model ID, format of parameters, model structure information, quantization method and parameters
- Proposal 6 (Sec 2.2): For MI-Option 3 procedure, model identification is achieved when the model ID is delivered together with the delivered model
- Proposal 7 (Sec 3.2): For model transfer/delivery where model is trained at UE side or neutral site, assume Case y as the baseline
- Proposal 8 (Sec 4): Lower the priority of UE-side training data collection discussion at RAN1 since content for use cases already provided in Rel-18 LS reply and UE data collection mechanisms are out of RAN1 scope