R1-2410778
discussion
Summary #4 for other aspects of AI/ML model and data
From OPPO
Summary
This 3GPP RAN1 technical document from OPPO summarizes discussions on AI/ML model and data aspects for NR air interface, containing approximately 30 proposals across model identification, training data collection, and model transfer/delivery topics. The document focuses on two-sided model scenarios and establishes frameworks for model identification procedures, dataset transfer mechanisms, and standardized model structures for Case z4 model transfer.
Position
OPPO, as the moderator, is advocating FOR a unified approach to AI/ML model identification that supports both functionality-based and model-ID-based operations, with network-assigned model IDs preferred for consistency. They push FOR standardized reference models (MI-Option4) and model transfer Case z4 with 3GPP-specified model structures rather than offline vendor coordination. OPPO is AGAINST overly complex model identification schemes that burden UE implementation and favors transparent network implementation decisions for dataset construction.
Key proposals
- Proposal 2.1A (Sec 5.1): For study of MI-Option2 (model identification with dataset transfer) for two-sided model, ID-X can be used for pairing the UE-part and the NW-part of a two-sided model
- Proposal 2.2 (Sec 1.3): Regarding MI-Option2 for two-sided model, at least the following information is transmitted with the dataset from network to UE: input data, labels, format/type information, dataset size, and validation/testing related info
- Proposal 2.4 (Sec 1.3): Regarding MI-Option2 for two-sided model, further study whether/what/how some information (e.g., model backbone) of NW-part should be transmitted along with the dataset
- Proposal 2.6 (Sec 1.3): Regarding MI-Option4 for two-sided model, study cases where standardized reference model is UE part, NW part, or both parts of two-side model
- Proposal 2.7 (Sec 1.3): If multiple reference models are standardized in MI-Option4, a model ID can be pre-defined for each reference model with pre-defined model IDs for identification
- Proposal 4.1 (Sec 3.3): RAN1 observations on aligning known model structures for Case z4: Opt.1 (3GPP specified) vs Opt.2 (offline alignment between NW and UE)
- Proposal 4.2 (Sec 3.3): Regarding study of model transfer/delivery Case z4, RAN1 focuses on the option with standardized known model structures (Opt.1)
- Proposal 4.3 (Sec 3.3): For Case z4 with standardized known model structures, study prioritizes standardized model structure of UE-sided model / UE part of two-sided model
- Proposal 4.4 (Sec 3.3): For Case z4 study, further study options for open format including reusing existing formats (ONNX), defining new 3GPP format (ASN.1), or reusing SA2 mechanisms
- Proposal 4.5B (Sec 3.3): For Case z4 model delivery/transfer used for inference, candidate solutions include UE signaling readiness or assuming readiness from minimum applicable time
- Proposal 4.6 (Sec 3.3): For Case z4 used for training/re-training/finetuning at UE-side, UE sends signaling to notify that AI/ML operations compatible to the transferred parameters are ready
- Proposal 4.7 (Sec 5.1): Study relationship of model ID, first indication, and second indication for Case z4: model ID as combination, second indication as model ID, or separate network-assigned model ID
- Proposal 4.8 (Sec 3.3): For standardized known model structures in Case z4, consider transformer, CNN, and LSTM backbones as starting point
- Proposal 2.5 (Sec 5.1): For MI-Option2, whether dataset includes data from one cell or multiple cells is up to network implementation and transparent to UE/UE-side
- Proposal 4.5B (Sec 5.2): For Case z4 model delivery/transfer used for inference, options include UE signaling readiness or assuming readiness from minimum applicable time after completion
Revision chain
2 versions in this meeting · oldest first
-
Summary #3 for other aspects of AI/ML model and data
-
R1-2410778 ← you are here discussion noted Final