R1-2407696
discussion
Discussion on AIML for CSI compression
From Spreadtrum
Summary
Spreadtrum presents evaluation results for AI-based CSI Spatial-Temporal-Frequency (S-T-F) compression showing superior performance over Rel-16 eType II codebook, and provides 5 proposals addressing inter-vendor training collaboration, CQI determination, historical CSI handling, and performance monitoring for AI/ML-based CSI compression use cases.
Position
Spreadtrum strongly advocates FOR AI-based CSI Spatial-Temporal-Frequency (S-T-F) compression demonstrating it achieves over double the SGCS gain compared to spatial-frequency compression alone, and AGAINST option 5a for inter-vendor collaboration due to increased complexity. They push FOR prioritizing options that only transfer CSI generation parts to UE to protect NW proprietary information, and advocate FOR Option 1b for CQI determination with realistic channel measurement and adjustment rather than simpler approaches.
Key proposals
- Proposal 1 (Sec Inter-vendor training collaboration): Deprioritize option 5a for inter-vendor training collaboration
- Proposal 2 (Sec Inter-vendor training collaboration): Options that only transfer CSI generation part to UE can be considered with high priority
- Proposal 3 (Sec CQI determination): For the study of CQI determination in inference, support Option 1b (CQI is calculated based on target CSI with realistic channel measurement and potential adjustment)
- Proposal 4 (Sec Handling of misalignment of historical CSI): Consider to report historical CSI information via NW-triggered signaling when UCI missing or UCI dropping
- Proposal 5 (Sec Monitoring): For performance monitoring, support NW side monitoring based on the ground-truth CSI reported by UE and UE side monitoring based on the recovery CSI indicated by NW