R1-2410195
discussion
Study on CSI compression
From LG Electronics
Summary
LG Electronics proposes 9 technical proposals for improving AI/ML CSI compression performance in NR, focusing on temporal/spatial/frequency domain compression, inter-vendor training collaboration, and addressing practical implementation challenges like UCI feedback issues and rank adaptation.
Position
LG Electronics strongly advocates FOR extending CSI compression to temporal/spatial/frequency (TSF) domain to leverage past channel information correlation, supporting dataset exchange method (Option 4-1) over model parameter exchange (Option 3a-1) for inter-vendor collaboration due to lower signaling overhead (100M-150M bits vs 192M-256M bits), and prioritizing Option 1 for CQI determination. They push AGAINST joint compression/prediction approaches that lack proper performance monitoring granularity and advocate FOR two-step performance monitoring to distinguish between model issues and past CSI quality problems.
Key proposals
- Proposal 1 (Sec 2.1.1): Study methods/mechanisms to manage the similarity/synchronization of accumulated past CSI at UE-side and/or NW-side for TSF-domain CSI compression
- Proposal 2 (Sec 2.1.1): Discuss the format of past CSI information and how to report it at least for performance monitoring perspective in TSF-domain CSI compression
- Proposal 3 (Sec 2.1.2): Consider performance monitoring method on joint CSI compression and prediction by adapting the operation on the AI/ML model between CSI compression and prediction for Case 3 and 4
- Proposal 4 (Sec 2.1.3): Consider two-step performance monitoring to check performance degradation origin and report past CSI information via NW-triggered signaling when UCI missing or dropping
- Proposal 5 (Sec 2.1.4): Consider the method on rank adaptation based on the availability check of layer(s) for a given RI
- Proposal 6 (Sec 2.1.5): For CQI determination in CSI compression using two-sided model, prioritize Option 1 and consider Option 2a utilizing AI/ML model complexity reduction method
- Proposal 7 (Sec 2.2): Consider assistant information regarding relationship between target CSI and reconstructed target CSI to reduce/alleviate signaling overhead and model alignment issue for Option 4-1
- Proposal 8 (Sec 2.2): Study model complexity reduction methods like knowledge distillation to further reduce CSI training/signaling complexity for Type 3 training collaboration
- Proposal 9 (Sec 2.2): Support Option 4-1 for Direction A for addressing inter-vendor training collaboration when considering down-selection between Option 4-1 and Option 3a-1