R1-2410044
discussion
On other aspects of AI/ML model and data
From InterDigital
Summary
InterDigital's contribution addresses AI/ML model identification for two-sided models, data collection for training, and model transfer/delivery aspects for NR air interface. The document contains 14 proposals and 9 observations across model identification challenges, CSI enhancement requirements, positioning use cases, and beam management optimization.
Position
InterDigital advocates for deferring complex model identification discussions until RAN4 makes progress, emphasizing that dataset-only exchange is insufficient for two-sided model compatibility. They push for LMF-centralized ground truth quality control in positioning, enhanced beam reporting up to 64 RSRP values with compression mechanisms, and against standardized model transfer for UE-side functionality-based LCM models, preferring specification-transparent delivery instead.
Key proposals
- Proposal 1 (Sec 2.1): Defer discussions on MI-Option4 until more progress is made in RAN4 (on RAN4-Option3, or RAN4-Option4)
- Proposal 3 (Sec 2.1): Defer further discussions on MI-Option2 until the conclusion of the evaluation effort on NW-side/UE-side additional conditions
- Proposal 4 (Sec 2.2.1): For model input and ground truth for CSI prediction model training dataset, the collected data could include the measured CSI during the observation and the prediction window
- Proposal 6 (Sec 2.2.1): Quality indicators for the CSI prediction model training dataset could include at least the RSRP and TDCP
- Proposal 8 (Sec 2.2.2.3): For case 1 for positioning, support LMF to forward location information of PRUs, measurements made by PRUs and ground truth label quality indicator with the PRU location to a target UE
- Proposal 10 (Sec 2.2.2.4): The LMF is the only entity that can generate a ground truth label quality indicator associated with location information of UE or PRU
- Proposal 11 (Sec 2.2.3): For UE side model, support a common procedure to measure whole Set A over multiple time instances for both BM-Case 1 and BM-Case 2
- Proposal 12 (Sec 2.2.3): For gNB side model, support enhanced UE reporting to report up to 64 RSRP values for whole Set A over multiple time instances
- Proposal 13 (Sec 2.2.3): Support beam reporting compression mechanism for training to reduce overhead by using RSRPs in neighboring beams in spatial domain and RSRPs within a same beam in temporal domain
- Proposal 14 (Sec 2.3): Model transfer for UE-side models with functionality-based LCM is not supported and 3GPP specification transparent model delivery is only considered