R1-2410104
discussion
Additional study on AI/ML-based CSI compression
From OPPO
Summary
OPPO presents analysis on inter-vendor collaboration approaches for AI/ML-based CSI compression, evaluating multiple directions and options for standardization. The document contains 17 proposals and 5 observations covering reference model standardization, parameter/dataset exchange methods, and data collection requirements.
Position
OPPO advocates FOR deprioritizing Direction C (fully standardized reference models) in Rel-19 due to high standardization effort and compatibility concerns, while pushing FOR prioritizing Case 0 model structure work and encoder-only specification in Direction C. They strongly support addressing data distribution mismatch issues through UE-side data reporting to enable NW-side fine-tuning, and advocate for proper model compression methods to address overhead concerns in over-the-air exchanges.
Key proposals
- Proposal 1 (Sec 2.1): Regarding reference model in Direction C, specify encoder only
- Proposal 6 (Sec 2.1): Deprioritize Direction C (Option 1) in Rel-19
- Proposal 7 (Sec 2.2): Regarding the standardization of model structure in Option 3, suggest to prioritize the work on Case 0. Case 2 and Case 3 can be further considered after model structure in Case 0 has been specified
- Proposal 8 (Sec 2.3): Regarding [issue 1] for Direction A (option 3a-1), additional information can include: Performance target, Dataset and dataset related information, Other necessary additional information is not precluded
- Proposal 9 (Sec 2.3): Regarding [issue 3] for Direction A and Direction B, overhead concern is more critical for over-the-air exchange compared to offline exchange, mainly from two aspects: Parameters exchange can be alleviated by proper model compression methods, Dataset exchange can be alleviated by target CSI label quantization methods
- Proposal 10 (Sec 2.3): Regarding [issue 4] for Direction A, [issue 6] for Direction B, performance impacts due to data distribution mismatch can be addressed by: UE reports UE-side data to NW, and NW performs encoder fine-tuning
- Proposal 11 (Sec 2.4): Regarding [issue 5] for Direction B, the feasibility should be considered from following two aspects: Performance aspect: feasible to train multiple encoders for different UEs, UE capability aspects: feasible if supporting standardizing multiple encoder structures
- Proposal 12 (Sec 2.6): RAN1 further study how to standardize data / dataset format in Option 4
- Proposal 13 (Sec 2.6): Regarding [Issue 1] for Direction A (Option 4-1), additional information can include: Performance target, Dataset related information, Other necessary additional information is not precluded
- Proposal 15 (Sec 2.6): Regarding [issue 4] for Direction A (Option 4-1), performance impacts due to data distribution mismatch can be addressed by: UE reports UE-side data to NW, and NW performs inference, and then exchange the CSI feedback information corresponding to UE-side data to UE
- Proposal 16 (Sec 2.7): Regarding the CSI-RS configuration for UE-side and NW-side data collection, at least the follow aspects should be considered: Data collection specific CSI-RS or not, Cell-specific or UE specific CSI-RS, Trade-off between performance and overhead
- Proposal 17 (Sec 2.7): Regarding the data collection for CSI compression, cell/site/scenario related condition information and addition condition information should be considered during the data collection stage