R1-2407749
discussion
Other aspects of AI/ML model and data
Summary
This 3GPP RAN1 document from Tejas Networks discusses AI/ML model identification and life cycle management for NR air interface, focusing on associated IDs for ensuring consistency between training and inference phases. The document contains 6 proposals and 4 observations covering model identification procedures, mapping between associated IDs and model IDs, and model transfer/delivery aspects.
Position
Tejas Networks advocates for UE-driven model ID assignment and supports maintaining current model transfer/delivery case prioritizations. They push FOR flexible many-to-one mapping between associated IDs and model IDs to enable generalized models, and FOR using existing RRC message structure elements like CSI-ResourceConfig for associated ID computation. They are AGAINST further deprioritizing model transfer cases and support keeping the current framework intact.
Key proposals
- Proposal 1 (Sec 3): We propose that the use-case specific data collection-related configurations and area or cite-specific information are mapped to associated ID(s)
- Proposal 2 (Sec 3): We claim that UE should assign a model ID and reports back to NW
- Proposal 3 (Sec 4): In many-to-one mapping of associated ID(s) to model ID, a single model is trained by a UE after collecting many associated ID(s), which may correspond to different NW side configurations or conditions
- Proposal 4 (Sec 4): In our view associated ID is computed based on PLMN ID, which uniquely identifies a mobile network geographically among others and several other IDs which had already existed in the legacy RRC message structure per use case
- Proposal 5 (Sec 4): We propose one or many of the IDs specified in IE CSI-ResouceConfig is valid for the computation of associated ID for both CSI compression and BM use cases
- Proposal 6 (Sec 6): We propose that the progress made so far on deprioritizing some model transfer or delivery cases is sufficient, and there is no need to further deprioritize the remaining cases
- Observation 1 (Sec 3): NW side additional conditions are signaled as associated ID(s)
- Observation 2 (Sec 3): There is a one-to-one mapping between associated ID(s) to data sets(s)
- Observation 3 (Sec 4): Mapping between associated ID(s) to model ID(s) take any of the following mappings: one-to-one, one-to-many, many-to-one, and many-to-many map
- Observation 4 (Sec 5): Multiple associated ID(s) may correspond to a single data set, and UE may train a model which is more generalized and robust enough to infer even in the case of moving to multiple cell(s) for which no data has been collected yet