R1-2407617
discussion
Discussion of additional study on AI/ML for NR air interface for CSI compression
From FUTUREWEI
Summary
Futurewei contributes to AI/ML for NR air interface CSI compression, discussing inter-vendor training collaboration options and providing temporal-domain CSI compression evaluation considering UCI loss. The document contains 10 proposals and 7 observations covering collaboration directions, quantization impacts, performance monitoring, and CQI determination.
Position
Futurewei advocates FOR using Rel-16 eType II codebook with new/enhanced parameters for data collection and monitoring to achieve better performance, even with additional overhead concerns. They push FOR comprehensive analysis of inter-vendor collaboration options through detailed tables showing issue applicability, and support using upper layer signaling to reduce air-interface overhead for less latency-sensitive options. They demonstrate strong support for temporal CSI compression (Case 2) showing significant performance gains over legacy approaches even with UCI loss.
Key proposals
- Proposal 1 (Sec 2.1.1): For issues identified for Direction A related to on-device operation and UE side offline engineering, consider the analysis captured in Table 2.1.1-1 and the corresponding notes
- Proposal 2 (Sec 2.1.2): For issues identified for Direction B related to on-device operation and UE side offline engineering, consider the analysis captured in Table 2.1.2-1 and the corresponding notes
- Proposal 3 (Sec 2.1.3): For issues identified for Direction C related to on-device operation and UE side offline engineering, consider the analysis captured in Table 2.1.3-1 and the corresponding notes
- Proposal 4 (Sec 2.2): Among the options to alleviate/resolve inter-vendor training collaboration issues, consider using standardized signalling in upper layers (case z4) as one delivery option to save air-interface overhead at least for Option 3a/4/5a
- Proposal 5 (Sec 2.3): For AI/ML-based CSI compression using two-sided model, further study potential specification impact related to quantization of CSI feedback including vector quantization codebook exchange, segmentation information, scalar quantization granularity configuration, and quantization dictionary exchange
- Proposal 6 (Sec 2.4.1): 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 7 (Sec 2.4.2): In AI/ML-based CSI compression using two-sided model, for NW-side monitoring based on target CSI reported by UE, consider adopting Rel-16 eType II CB based quantization with new parameters to achieve better monitoring performance
- Proposal 8 (Sec 2.4.2): In AI/ML-based CSI compression using two-sided model, for UE-side monitoring, if CSI reconstruction model or reference model is available at UE, support monitoring based on output of CSI reconstruction model at UE (Case 2-1)
- Proposal 9 (Sec 2.4.2): In AI/ML-based CSI compression using two-sided model, for UE-side monitoring, further study LCM complexity associated with using proxy reconstruction model first before discussing whether to support proxy reconstruction model at UE side for performance monitoring
- Proposal 10 (Sec 2.4.3): In AI/ML-based CSI compression using two-sided model, for CQI determination, if CSI reconstruction model or reference model is available at UE, adopt Option 2a to determine CQI at UE