R1-2409398
discussion
Discussion on AI/ML for CSI compression
From Huawei
Huawei's prior position on
9.1.4.1
at
RAN1#118bis
· AI-synthesized, paraphrased
verify sources →
Supports Direction A (dataset/model sharing for UE-side offline engineering) over Direction B, and prioritizes Option 4 dataset sharing with simpler sub-options due to significantly lower air-interface overhead.
Summary
This Huawei contribution discusses inter-vendor training collaboration for AI/ML-based CSI compression in NR Rel-19, focusing on Direction A (dataset/parameter exchange) and temporal domain cases. It presents 17 proposals and 11 observations covering training methods, overhead alleviation, specification impacts for temporal CSI, data collection formats, and monitoring mechanisms.
Position
Huawei argues that for Direction A inter-vendor collaboration, sharing datasets (Option 4-1) incurs less proprietary risk than sharing model parameters (Option 3a-1/3b). They propose that model backbone information is unnecessary for Option 4-1, allowing UE autonomy in structure selection guided by NW performance targets. Regarding temporal domain CSI compression, Huawei presents technical cases showing that UCI missing causes significant performance loss due to accumulated CSI misalignment, proposing reset mechanisms to mitigate this. For NW-side data collection, they prioritize precoding matrices over channel matrices and advocate for Rel-16 eType II codebook with enhanced parameters (e.g., L=8, 10, 12) to balance overhead and accuracy. They oppose specifying UE-side proxy models for monitoring, citing imbalanced generalization and excessive LCM burden, instead supporting NW-side ground-truth CSI monitoring and UE-side precoded RS monitoring.
Key proposals
- Proposal 1 (Sec 2.2.1): Study performance targets for Direction A, including end-to-end, encoder-only, and decoder-only metrics, noting targets are only applicable to NW-shared testing datasets.
- Proposal 2 (Sec 2.2.2): Argue that model backbone information is not needed for Option 4-1, as UEs can autonomously select structures based on NW-indicated performance targets.
- Proposal 3 (Sec 2.2.3): Propose sharing at least Target CSI from NW to UE for Option 3a-1 to ensure dataset alignment and avoid generalization problems.
- Proposal 4 (Sec 2.3): Question the necessity of resolving data distribution mismatch issues, arguing NW-side diverse datasets and timely updates can alleviate mismatch.
- Proposal 5 (Sec 2.4.1): Propose splitting overall datasets into subsets for air-interface delivery to relieve overhead, with re-combination at UE-side OTT servers.
- Proposal 6 (Sec 3.1): Propose down-selecting factors for standardized model structure in Option 3, such as input type, rank>1 options, and quantization methods.
- Proposal 7 (Sec 3.2): Propose studying standardized dataset format for Option 4, including data sample format, construction aspects, scalability, and target performance.
- Proposal 8 (Sec 4.1): Propose considering methods to handle misalignment of accumulated CSI between NW and UE models due to UCI missing in temporal Case 2.
- Proposal 9 (Sec 4.2): Propose studying spec impact for temporal Case 3, specifically data collection and monitoring for separate prediction/compression, and data collection/inference/monitoring for joint prediction/compression.
- Proposal 10 (Sec 5.1): Propose confirming NW-side data collection feasibility, prioritizing precoding matrix over channel matrix, and using Rel-16 eType II CB with new parameters (e.g., L=8,10,12).
- Proposal 11 (Sec 5.2.1): Propose NW-side monitoring based on ground-truth CSI in Rel-19, preferring eT2-like high-resolution codebooks and SGCS as the intermediate KPI.
- Proposal 12 (Sec 5.2.2): Argue there is no strong motivation for specifying UE-side proxy models (Case 2-1/2-2) for monitoring due to generalization and LCM burdens.
- Proposal 13 (Sec 5.2.3): Propose UE-side monitoring based on precoded RS transmitted from NW, studying KPIs calculated from these RS.
- Proposal 14 (Sec 5.3.1): Propose studying quantization alignment using standardized schemes, including dictionary configuration for VQ and granularity for SQ.
- Proposal 17 (Sec 5.3.4): Propose prioritizing Rank>1 options 3-1, 3-2, and 2-1, discussing whether multiple models are handled separately or as a single model for LCM.