Discussion on AI/ML for CSI compression

Summary

This Huawei document analyzes AI/ML for CSI compression in Release 19, focusing on inter-vendor training collaboration, temporal domain extensions, and remaining specification issues from Release 18. The document contains 17 proposals and 10 observations covering training methods, overhead concerns, monitoring approaches, and inference aspects.

Position

Key proposals

• Proposal 1 (Sec 2.2.1): For additional information of Direction A (Option 3a/5a/4), study the performance target including end-to-end performance metric and/or encoder only performance metric, type of performance metric, and testing dataset used for metric calculation
• Proposal 3 (Sec 2.2.2): Regarding the data distribution mismatch issue of Direction A/B, the necessity and feasibility to resolve it may need careful justification, considering NW side can collect sufficiently diverse training dataset and timely dataset updates
• Proposal 4 (Sec 2.2.3): For dataset sharing for Option 3a/4/5a over air-interface, study solution to relieve overhead by splitting overall dataset into many subsets with limited data samples associated with common dataset ID for UE side re-combination
• Proposal 5 (Sec 2.2.4): For further study of Direction A/B, consider Option 3a-3/4-3/5a-3 with lower priority due to larger overhead and higher risk of proprietary disclosure with no clear benefit on performance
• Proposal 6 (Sec 2.3.1): For studying standardized model structure of inter-vendor collaboration Option 3, strive to perform down selection to factors so that single model structure along with pre/post processing approach is aligned before starting evaluation
• Proposal 8 (Sec 3.1): For additional potential spec impact of temporal domain CSI compression Case 2, consider methods to handle misalignment of accumulated CSI between NW part model and UE part model due to UCI missing
• Proposal 9 (Sec 3.2): For additional potential spec impact of temporal domain CSI compression Case 3, potential spec impact may be needed for data collection, inference, and monitoring for separate and joint prediction and compression
• Proposal 10 (Sec 4.1): For NW side data collection, prioritize precoding matrix over channel matrix, prioritize Rel-16 eType II CB based quantization with new parameters L=8,10,12; pv=0.8,0.9,0.95; reference amplitude=6,8 bits
• Proposal 11 (Sec 4.2.1): For NW-side monitoring, consider ground-truth CSI based monitoring with eT2-like high-resolution codebook for reporting format and SGCS for type of intermediate KPI with higher priority
• Proposal 12 (Sec 4.2.2): There is no strong motivation for specifying the UE side proxy model (Case 2-1/2-2) for monitoring due to imbalanced generalization performance and additional LCM burden
• Proposal 13 (Sec 4.2.3): For UE-side monitoring, consider precoded RS transmitted from NW based on output of CSI reconstruction model based monitoring in Rel-19
• Proposal 14 (Sec 4.3.1): For quantization methods of CSI report, further study potential specification impact on quantization alignment using standardized quantization scheme for vector and scalar quantization
• Proposal 15 (Sec 4.3.2): For study of CQI determination in inference, consider Option 1 where CQI is NOT calculated based on output of CSI reconstruction part from realistic channel estimation as starting point
• Proposal 16 (Sec 4.3.3): For CSI report in inference, additionally study AI/ML specific aspects including CSI priority rules, CSI processing unit considering UE part model complexity differences, and CSI mapping
• Proposal 17 (Sec 4.3.4): For Rank>1 options in inference, further study Option 3-1 (layer common and rank common), Option 3-2 (layer common and rank specific) and Option 2-1 (layer specific and rank common) with higher priority

Your notes