R1-2410724
discussion
Final summary of Additional study on AI/ML for NR air interface: CSI compression
From Qualcomm
Summary
This 3GPP RAN1 document (R1-2410724) from Qualcomm serves as the final meeting summary for AI/ML-based CSI compression studies in Release 19, containing over 200 proposals from multiple companies across temporal domain aspects, inter-vendor collaboration, monitoring, and data collection. The document concludes the study phase with key agreements on prioritizing Case 2 and Case 3 temporal compression, adopting Direction A (dataset/parameter exchange) and Direction C (fully standardized models) for inter-vendor collaboration, and establishing requirements for NW-side data collection using enhanced eType-II codebooks.
Position
Qualcomm, as the document moderator, advocates for a pragmatic approach prioritizing Direction A Option 4-1 (dataset sharing) over Option 3a-1 due to lower specification complexity, while supporting Direction C as a minimum performance baseline. They push for NW-side target CSI sharing to address data distribution mismatch and emphasize the need for enhanced eType-II codebooks for effective monitoring, positioning against premature down-selection of options without proper feasibility analysis from RAN2.
Key proposals
- Proposal 21b (Sec 1): Under 10% UCI loss, Case 2 shows small to large performance drop, with signaling support including NW-signaling to reset historical CSI information and NW-triggered CSI retransmission to mitigate UCI loss impact
- Proposal 22c (Sec 2): For temporal domain Case 3, study both Separate Prediction and Compression (SPC) and Joint Prediction and Compression (JPC) with different options for training data collection and monitoring targets derived from predicted vs measured CSI
- Proposal 41a (Sec 3): Send LS to RAN2 to study feasibility of parameter/dataset exchange using over-the-air and other signaling approaches for Direction A inter-vendor collaboration
- Proposal 45b (Sec 3): Fully specified models in Direction C may have limited field performance compared to Direction A, but can be improved if either/both sides train using field data, though still worse than achievable Direction A performance
- Proposal 49a (Sec 3): For Direction A and C, at least support Direction A Option 4-1 (dataset sharing), with feasibility of Option 3a-1 and Direction C contingent on scalable model structure specification
- Proposal 47b (Sec 4): For standardized model structure, adopt Transformer backbone for Case 0 spatial-frequency input, with Case 2/3 adaptations including Conv-LSTM, LSTM, and latent adaptation options
- Proposal 51b (Sec 4): Confirm necessity of ground-truth reporting for NW-side data collection using Rel-16/18 eType2 codebooks with potential enhancements, including configuration of rank/layer and additional sample information
- Proposal 52b (Sec 4): Confirm feasibility of UE-side data collection with NW configuration or UE request, including RS configuration and temporal aspects for Cases 2/3
- Proposal 61a (Sec 5): Confirm necessity of NW-side monitoring via eType2 codebook contingent on UE capability, and necessity of additional monitoring options for UEs not supporting concurrent eType2 and two-sided CSI compression
- Proposal 64b (Sec 5): For NW-first training in options 3/4/5, root cause identification achieved via ground-truth CSI reporting where NW runs inference using reference CSI generation part compared to UE side part
- Proposal 81a (Sec 6): Study performance-complexity trade-off comparing different AI/ML models and precoder representations in spatial-frequency vs angle-delay domains considering FLOP complexity and performance metrics
- Proposal 43b (Sec 3): For Option 3a-1, NW-side target CSI sharing needed for UE-side encoder performance validation/testing, with FFS on need for training to address NW/UE data distribution mismatch