R1-2410260
discussion
Discussion on AI/ML for CSI compression
From ETRI
Summary
ETRI presents comprehensive views on AI/ML-based CSI compression for NR air interface Release-19 further study, addressing temporal domain aspects and inter-vendor training collaboration. The document contains 8 proposals and 13 observations covering specification impacts, training collaboration directions, and performance monitoring approaches.
Position
ETRI advocates FOR supporting both Direction A (UE-side offline engineering) and Direction B (direct parameter sharing) inter-vendor training collaboration approaches, emphasizing the importance of maximizing reuse of existing CSI-RS specifications for temporal domain aspects. They push AGAINST Direction C due to severe performance degradation (37.2% SGCS loss) and advocate for UE-side performance monitoring (Case 2-1) over network-side alternatives due to complexity and overhead considerations.
Key proposals
- Proposal 1 (Sec 2.1.1): For AI/ML-based CSI compression using two-sided model, when UE and/or NW uses past CSI information, reuse the current specification on CSI-RS transmissions as much as possible
- Proposal 2 (Sec 2.1.2): For AI/ML-based CSI compression using two-sided model, when NW uses past CSI information, study method to detect and mitigate inconsistency of the availability of past CSI information between the UE and the NW
- Proposal 3 (Sec 2.1.3): For AI/ML-based CSI compression using two-sided model, when the target CSI is Future slot(s), study method to align whether prediction and compression occur in separate steps or simultaneously between UE and NW, and either UE-side or NW-side performance monitoring
- Proposal 4 (Sec 2.2.1): Regarding Direction A and B, support both directions, timely collected data on NW side can prevent performance degradation, sharing NW-side target CSI may not always be necessary, and Option 3a-1 is preferred over Option 4-1 due to lower overhead
- Proposal 5 (Sec 2.2.2): Regarding Direction C, only UE-side offline engineering to the specified decoder (Option 1-2) is feasible without inter-vendor training collaboration
- Proposal 6 (Sec 2.3.1): For AI/ML-based CSI compression using two-sided model, study functionality/model identification procedure for supporting both Direction A and B
- Proposal 7 (Sec 2.3.2): For dataset delivery for training collaboration type 3, consider the limited number of dataset samples to assess the feasibility of incorporating standardized signaling
- Proposal 8 (Sec 2.3.3): For the performance monitoring of CSI compression sub-use case using two-sided model, conclude that Case 2-1 is feasible with considerations of complexity, latency, and accuracy