R1-2409790
discussion
Discussion on AI based CSI compression
From Apple
Summary
Apple analyzes inter-vendor training collaboration options for AI/ML-based CSI compression, arguing that Option 3a-1 should be deprioritized in favor of Option 4-1 with assisted nominal decoder training to handle UE-side additional conditions. The document presents 11 proposals and 12 observations covering specification impacts for time-frequency-spatial domain CSI compression cases, performance monitoring mechanisms using precoded CSI-RS, and root cause detection strategies for data drift.
Position
Apple deprioritizes inter-vendor training collaboration option 3a-1, arguing it cannot handle UE side additional conditions as effectively as option 4-1 with Alt 1, which utilizes a nominal decoder. They propose that option 4-1 requires additional assisted information, including reference decoder model backbone and structure, to ensure the nominal decoder resembles the actual gNB decoder. For time-frequency-spatial domain CSI compression, Apple proposes enabling semi-persistent CSI reporting and DCI-based state reset for cases 2 and 4 to mitigate SGCS loss from UCI drop-induced state desynchronization. Regarding performance monitoring, they argue for using precoded CSI-RS to implicitly transmit output CSI to the UE, providing a low-overhead solution compared to extended parameter sets, and propose studying RLF/BFD-like mechanisms for UE-initiated reports. Finally, they propose using UL SRS for detecting channel statistics drift and immediate MDT for collecting root cause detection data when distribution mismatch is suspected.
Key proposals
- Proposal 1 (Inter-vendor training collaboration): Deprioritize training collaboration option 3a-1 due to similar performance to option 4-1 but larger specification work.
- Proposal 2 (Inter-vendor training collaboration): Define additional information for option 4-1 to assist UE in training a nominal decoder, including performance targets, dataset categorization assistance, and reference decoder structure.
- Proposal 3 (Potential specification impact for case 2): For time-frequency-spatial domain CSI compression cases 2 and 4, enable semi-persistent CSI reporting, DCI-based state reset, UCI retransmission for sync, and longer-term RI updates.
- Proposal 4 (Potential specification impact for case 2): Use options discussed for frequency-spatial domain CSI compression for NW and UE side performance monitoring in case 2.
- Proposal 5 (Potential specification impact for case 4): For case 4 performance monitoring, use CSI measurement in the prediction window as target CSI, with KPIs including both compression and prediction performance.
- Proposal 6 (Potential specification impact for case 3): For case 3 performance monitoring, use CSI measurement in the prediction window as target CSI, with KPIs including both compression and prediction performance.
- Proposal 7 (CSI report configuration): Study flexible CSI report configuration to support different time-frequency-spatial domain CSI compression cases.
- Proposal 8 (Remaining issue of spatial-frequency domain CSI compression): Further study RLF/BFD-like mechanisms for UE-initiated reports in two-sided model CSI compression.
- Proposal 9 (Root cause detection): Use UL SRS measurement at gNB to detect data drift due to channel statistics changes.
- Proposal 10 (Root cause detection): Trigger root cause detection data collection via immediate MDT with logged data when performance degradation is detected without observed channel statistics drift.
- Proposal 11 (Root cause detection): Further study root cause detection for data distribution mismatch due to UE side additional condition.