R1-2410469
discussion
Additional study on CSI compression
From Qualcomm
Summary
Qualcomm's comprehensive technical document on AI/ML-based CSI compression for two-sided models presents 25 proposals and 22 observations covering inter-vendor collaboration, model performance monitoring, inference aspects, and complexity reduction. The document advocates for dataset sharing (option 4-1) over parameter sharing, UE-side monitoring via SGCS estimators, and deprioritizing over-the-air signaling solutions.
Position
Qualcomm strongly advocates FOR dataset sharing (option 4-1) over encoder parameter sharing (3a-1), UE-side performance monitoring using SGCS estimators, and non-over-the-air solutions for model/dataset exchange. They are AGAINST over-the-air delivery of training data as unscalable, direction B (option 3b) due to feasibility concerns, and angle-delay domain approaches for small payload regimes. They push for RAN2 involvement in signaling solutions and emphasize the need for proprietary model accommodation while ensuring minimum performance through RAN4 models.
Key proposals
- Proposal 1 (Sec 2): Capture conclusions for direction A issues - SGCS requirements/guidance is beneficial, information to facilitate UE side data collection needed, overhead for dataset sharing needs study but less concern if not over-the-air
- Proposal 2 (Sec 2): Prioritize option 4-1 over 3a-1 with target CSI sharing unless additional spec impact of encoder parameter sharing and model structure standardization is justified
- Proposal 5 (Sec 2): Deprioritize direction B (option 3b) for addressing inter-vendor collaboration complexity in Rel-19 two-sided CSF study
- Proposal 8 (Sec 2): Prioritize dataset sharing (4-1) in Direction A, and using specified RAN4 model to ensure minimum performance in direction C
- Proposal 12 (Sec 2.2): Over-the-air delivery/transfer of training dataset, parameter, or reference model to support UE-side model training is not suitable or scalable solution
- Proposal 13 (Sec 2.2): RAN2 concludes that standard-based solutions (other than over Uu) for delivering dataset/parameter/reference model from network entity to UE-side training server is feasible for Rel-20
- Proposal 15 (Sec 2.3): For two-sided CSF use case, consider using pairing ID to keep consistency between training and inference, study specification impact in data collection RS request/configuration
- Proposal 18 (Sec 3): Consider two-phase mechanism for performance monitoring - Phase 1: identifying performance degradation, Phase 2: identifying root cause (UE side, NW side, data drift)
- Proposal 19 (Sec 3): Consider NW side monitoring with ground-truth reporting and UE side monitoring with Phase 1 via SGCS estimator, Phase 2 with ground-truth reporting
- Proposal 20 (Sec 4): Conclude CQI calculation option 2a can be employed with consideration of potentially higher timeline and cost of processing unit and memory
- Proposal 22 (Sec 4.2): Consider layer-common and rank common (Option 3-1) structure for CSI generation/reconstruction model for specified structures
- Proposal 23 (Sec 4.3): Study proprietary quantization codebook with standardized exchange signaling from NW side to UE side for inter-vendor collaboration direction A and B
- Proposal 24 (Sec 4.4): Complexity reduction by representing precoder in angle-delay domain needs further study considering SGCS performance, complexity reduction, and overhead bits for signaling SD/FD basis
- Proposal 10 (Sec 2): RAN2 should be involved to study feasibility of specified encoder parameter and/or dataset sharing
- Proposal 17 (Sec 3): RAN1 to perform evaluation study for monitoring decision accuracy in terms of false alarms and miss-detection with specific methodology for datasets A and B