R1-2410777
discussion
Summary #3 for other aspects of AI/ML model and data
From OPPO
Summary
This OPPO-moderated document (R1-2410777) presents a comprehensive summary for RAN1 agenda item 9.1.4.2 on AI/ML model and data aspects, containing 31 proposals and 15 observations covering model identification procedures, training data collection, and model transfer/delivery mechanisms for Release 19.
Position
OPPO, as the moderator, advocates FOR a unified LCM framework supporting both functionality-based and model ID-based operations, with network assignment of model IDs preferred for consistency. They push FOR studying model transfer Case z4 with standardized structures while being AGAINST premature discussions on aspects lacking sufficient progress, and favor reducing UE-side burden through multi-cell associated ID consistency assumptions.
Key proposals
- Proposal 2.2 (Sec 2): For MI-Option2 dataset transfer for two-sided models, transmit at least: input data corresponding to UE part input, labels corresponding to UE part output and their associations, format/type of input data and labels, dataset size, and validation/testing related info
- Proposal 2.4 (Sec 2): For MI-Option2 dataset transfer, further study whether/what/how information (e.g., model backbone) of NW-part of two-sided model should be transmitted along with the dataset
- Proposal 2.6 (Sec 2): For MI-Option4 standardized reference models for two-sided models, study at least three cases: Case-MI-4A (standardized UE part), Case-MI-4B (standardized NW part), Case-MI-4C (standardized both parts)
- Proposal 2.7 (Sec 2): For MI-Option4 with multiple reference models, pre-define model IDs for each reference model that can be identified by UE/network, with FFS on ID details and whether UE-developed models compatible with same reference need network identification
- Proposal 4.2 (Sec 4): For model transfer/delivery Case z4 necessity/benefit study, RAN1 focuses on option with standardized known model structures (Opt.1), noting offline alignment (Opt.2) is beyond RAN1 expertise scope
- Proposal 4.3 (Sec 4): For Case z4 with standardized known model structures for inference, prioritize studying standardization of UE-sided model/UE part of two-sided model structures
- Proposal 4.4 (Sec 4): For Case z4 study, investigate open format options including: reusing existing AI community formats (e.g., ONNX), defining new 3GPP format (including ASN.1), or reusing SA2 interoperability token mechanism
- Proposal 4.5B (Sec 4): For Case z4 model readiness determination when directly used for inference, consider either/combination of: UE signaling readiness notification, or model assumed ready after minimum application time (FFS whether time is specified or UE-reported)
- Proposal 4.6 (Sec 4): For Case z4 when used for training/re-training/fine-tuning at UE-side, UE sends signaling to notify network that UE supports AI/ML operations compatible to the transferred model parameters for inference readiness
- Proposal 4.8 (Sec 4): For standardized known model structures in Case z4 inference, consider transformer, CNN, and LSTM backbones as starting points
- Proposal 2.1A (Offline Agreement, Sec 2): For MI-Option2 study for two-sided models, ID-X can be used for pairing UE-part and NW-part of two-sided model, with FFS on other information needed for pairing
Revision chain
2 versions in this meeting · oldest first
-
R1-2410777 ← you are here discussion revised
-
Summary #4 for other aspects of AI/ML model and data