R1-2410051
discussion
Discussion on CSI compression with AI/ML
From Fujitsu
Summary
Fujitsu's comprehensive technical document analyzes CSI compression with AI/ML for NR air interface, presenting evaluation results for non-ideal UCI feedback and detailed discussions on inter-vendor training collaboration across three directions (A, B, C). The document contains 24 proposals and 21 observations covering topics from data distribution mismatch to performance monitoring approaches.
Position
Fujitsu strongly advocates FOR Direction A option 4-1 over 3a-1 due to similar performance with less specification effort, synthetic data training for reference models in Direction C, NW-side monitoring over UE-side proxy model monitoring due to reliability concerns, and codebook-based quantization approaches. They push AGAINST UE-side proxy model monitoring citing generalization and reliability issues, and against Option 2 for CQI determination due to complexity and delay concerns.
Key proposals
- Proposal 1 (Sec 3.1.1): To address performance degradation due to SVD phase normalization misalignment, consider Alt.1: specify phase normalization for reference encoder input, or Alt.2: introduce a phase correction layer for reference encoder input
- Proposal 7 (Sec 3.1.2): Considering no performance advantage without decoder related information and more specification effort required for option 3a-1, RAN1 to consider option 4-1 with high priority
- Proposal 9 (Sec 3.3.1): For Direction C, if synthetic data is used to train reference model, RAN1 to consider the standardized decoder with high priority
- Proposal 10 (Sec 3.3.2): For Direction C, RAN1 to consider training reference models with synthetic data
- Proposal 12 (Sec 4.1): For CSI compression using two-sided models, RAN1 to further discuss using codebook-like approach to report ground-truth CSI for AI/ML model training and monitoring, e.g. Rel-16 e-type II-like codebook with enhanced parameter values
- Proposal 13 (Sec 4.2): For Case 2, RAN1 to study how to address layer disorder issue for training data collection, model inference and performance monitoring
- Proposal 14 (Sec 4.3.1): Support both alternatives of precoding matrix for output-CSI-UE and input-CSI-NW: Alt 1: spatial-frequency/spatial-frequency-time domain for Case 2/Case 3, Alt 2: angular-delay/angular-delay-doppler domain projection
- Proposal 16 (Sec 4.3.3): For CSI compression using two-sided AI/ML models, deprioritize Option 2 proposed in RAN1 #112 for CQI determination based on CSI reconstruction part output from realistic channel estimation
- Proposal 19 (Sec 4.4.1): The feasibility, reliability, and generalization capability of UE-side AI/ML model performance monitoring using proxy model(s) should be evaluated and concluded before further specification impacts discussion
- Proposal 20 (Sec 4.4.2): For NW-side AI/ML model performance monitoring for CSI compression, RAN1 to prioritize study of codebook-based quantization method to obtain ground-truth CSI, including adding new parameter values to legacy codebook
- Proposal 23 (Sec 4.4.3): RAN1 to study procedures and signalling needed for follow-up actions after AI/ML model performance monitoring, including falling back to legacy codebook-based CSI reporting from AI/ML-based methods