R1-2410030
discussion
Discussion of CSI compression on AI/ML for NR air interface
From FUTUREWEI
FUTUREWEI's prior position on
9.1.4.1
at
RAN1#118bis
· AI-synthesized, paraphrased
verify sources →
Advocates for using Rel-16 eType II codebook with enhanced parameters for data collection and monitoring, and pushes for comprehensive analysis of inter-vendor collaboration options despite additional overhead concerns.
Summary
Futurewei presents a comprehensive analysis of AI/ML-based CSI compression for NR air interface, covering inter-vendor training collaboration options, specification impacts, and performance evaluation results. The document contains 10 formal proposals and 6 observations addressing parameter/model exchange methods, quantization impacts, data collection, performance monitoring, and CQI determination aspects.
Position
Futurewei advocates FOR using enhanced Rel-16 eType II codebook with new parameters for both data collection and monitoring to achieve better performance, despite overhead concerns. They push FOR differentiated delivery methods based on latency requirements (over-the-air for Direction B, upper layer signaling for Direction A) and support comprehensive quantization specification impact studies. They are AGAINST rushing into proxy model adoption without thorough LCM complexity analysis and advocate for careful performance evaluation before supporting precoded RS-based monitoring methods.
Key proposals
- Proposal 1 (Sec 2.1): If inter-vendor training collaboration option 3b in Direction B is supported, consider adopting over-the-air delivering method(s)
- Proposal 2 (Sec 2.1): If inter-vendor training collaboration option 3a-1 and/or 4-1 in Direction A are/is supported, consider adopting standardized signalling in upper layers or other offline method(s) as delivery options
- Proposal 3 (Sec 2.2): For AI/ML-based CSI compression using two-sided model, further study specification impact related to quantization of CSI feedback including vector quantization codebook exchange and scalar quantization configuration
- Proposal 4 (Sec 2.3.1): In AI/ML-based CSI compression using two-sided model, at least for NW-side ground-truth data collection for model training, consider adopting Rel-16 eType II CB based quantization with new parameters to achieve better performance
- Proposal 5 (Sec 2.3.2.1): In AI/ML-based CSI compression using two-sided model, for NW-side monitoring based on the target CSI reported by the UE, consider adopting Rel-16 eType II CB based quantization with new parameters
- Proposal 6 (Sec 2.3.2.2): For UE-side monitoring, if the actual or reference CSI reconstruction model is available at UE, support based on the output of the CSI reconstruction model at the UE (Case 2-1)
- Proposal 7 (Sec 2.3.2.2): For UE-side monitoring, if the actual or reference CSI reconstruction model is not available at UE, support based on the output indicated by NW via legacy eT2 codebook or eT2-like high-resolution codebook
- Proposal 8 (Sec 2.3.2.2): Further study the monitoring performance of precoded RS-based methods before discussing whether to support this option
- Proposal 9 (Sec 2.3.2.2): Further study at least the LCM complexity associated with using a proxy model first before discussing whether to support the use of proxy model at UE side
- Proposal 10 (Sec 2.4.1): For CQI determination, if the actual or reference CSI reconstruction model is available at UE, adopt Option 2a to determine CQI at UE