R1-2410721
discussion
Summary#3 of Additional study on AI/ML for NR air interface: CSI compression
From Qualcomm
Summary
This 3GPP RAN1 technical document (R1-2410721) from Qualcomm serves as the draft summary for AI/ML-based CSI compression study, containing over 100 company proposals across temporal domain aspects, inter-vendor collaboration, monitoring, and inference aspects. The document focuses on addressing UCI loss mitigation, standardizing model structures (particularly Transformer backbone), and resolving inter-vendor training collaboration complexity through multiple directional approaches.
Position
Qualcomm, as the moderator, advocates FOR: (1) Prioritizing Direction A Option 4-1 (dataset exchange) over Option 3a-1 due to lower specification complexity while maintaining performance, (2) Supporting Direction C for minimum performance guarantee using RAN4-compatible models, (3) Transformer backbone standardization for Case 0 with scalable model structures, and (4) NW-side monitoring as primary approach with ground-truth CSI reporting. They push AGAINST: (5) Direction B due to feasibility concerns with UE-common/specific encoders, (6) Over-the-air signaling for large dataset/parameter exchange, and (7) Joint prediction and compression approaches in favor of separate processing for reduced complexity.
Key proposals
- Proposal 21a (Sec 2): Under 10% UCI loss, Case 2 shows small to large performance drop. Consider NW-signaling to reset historical CSI information and NW-triggered CSI retransmission for mitigation
- Proposal 22a (Sec 2): For temporal domain Case 3, take separate prediction and compression (SPC) as baseline over joint prediction and compression (JPC), studying LCM aspects and specification impacts
- Proposal 49a (Sec 3): Support Direction A Option 4-1 at minimum, with UE assuming NW may deploy specified reference model (Direction C) or different model via inter-vendor collaboration
- Proposal 47b (Sec 3): Adopt Transformer as backbone structure for Case 0 spatial-frequency domain input, with Case 2/3 adaptations through input/output or latent adaptation methods
- Proposal 51b (Sec 4): Confirm necessity of ground-truth reporting for NW-side data collection using Rel-16/18 eType2 codebook format, with FFS on parameter enhancements
- Proposal 52b (Sec 4): Confirm necessity of UE-side data collection with NW configuration/UE request, including temporal aspects configuration and NW-side additional condition IDs
- Proposal 61a (Sec 5): Confirm necessity of NW-side monitoring based on target CSI reported via eType2 codebook, contingent on UE capability, plus additional monitoring for non-supporting UEs
- Proposal 62a (Sec 5): Evaluate monitoring decision performance using false alarm and missed detection probability as KPIs, with N samples large enough for accurate decisions
- Proposal 64a (Sec 5): For NW-first training in inter-vendor collaboration options 3/4/5, conclude root cause identification achievable through ground-truth CSI reporting
- Proposal 81a (Sec 7): Study performance-complexity trade-off comparing different AI/ML models and precoder representations (spatial-frequency vs angle-delay domains)
- Agreement (Historical): Prioritize Case 2 and Case 3 for temporal domain aspects of AI/ML-based CSI compression using two-sided model
- Agreement (Historical): For Direction A, performance impact due to NW/UE data distribution mismatch can be addressed by timely NW-side data collection and UE-side nominal decoder training
- Conclusion 42a (Sec 3): Direction A performance impact due to NW/UE data distribution mismatch addressed by timely NW-side data collection and training nominal decoder using NW-side target CSI
- Proposal 45b (Sec 3): Fully specified models in Direction C have limited field performance vs Direction A, but can be improved through field data training by either/both sides
- Discussion 43a (Sec 3): Preference elaboration for Direction A sub-options, with companies split between Option 4-1 (dataset exchange) and Option 3a-1 (parameter exchange with/without target CSI)