R1-2410723
discussion
Summary#5 of Additional study on AI/ML for NR air interface: CSI compression
From Qualcomm
Summary
This 3GPP RAN1 document from Qualcomm presents a comprehensive summary of AI/ML-based CSI compression study results with over 100 proposals from multiple companies covering temporal domain aspects, inter-vendor training collaboration, monitoring, and data collection. The document captures extensive evaluation results on UCI loss impact and separate vs. joint prediction and compression for Case 3, with conclusions on multiple key technical directions.
Position
Qualcomm, as the moderator, advocates for a balanced approach prioritizing Direction A Option 4-1 (dataset exchange) for inter-vendor collaboration while maintaining Direction C (fully specified models) as a minimum performance baseline. They push against premature down-selection of options, emphasizing the need for comprehensive feasibility studies including RAN2 liaison for signaling approaches, and support flexible solutions accommodating both UE-side offline engineering and standardized reference models.
Key proposals
- Proposal 21b (Sec 9.1.4.1): Conclude that under 10% UCI loss, Case 2 shows small to large performance drop, and signaling support to mitigate UCI loss can improve performance through NW-signaling to reset historical CSI information and NW-triggered CSI retransmission
- Proposal 22c (Sec 9.1.4.1): For temporal domain Case 3, take separate prediction and compression (SPC) as baseline and study LCM aspects, with options for training data collection using predicted vs measured CSI and separate monitoring approaches
- Proposal 49a (Sec 9.1.4.1): Support Direction A Option 4-1 at minimum for inter-vendor collaboration, with UE assuming NW may deploy specified reference model (Direction C) or different model via inter-vendor collaboration
- Proposal 47b (Sec 9.1.4.1): For standardized model structure, adopt Transformer as backbone for Case 0 spatial-frequency domain, with Case 2/3 reusing Case 0 structure plus adaptations like Conv-LSTM, LSTM, or latent adaptation
- Proposal 45b (Sec 9.1.4.1): Conclude 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 Direction A performance
- Proposal 51b (Sec 9.1.4.1): Confirm necessity of ground-truth reporting for NW-side data collection using Rel-16/18 eType2 codebook format, with FFS on enhancement needs and mechanisms outside L1-signaling
- Proposal 52b (Sec 9.1.4.1): Confirm necessity of UE-side data collection with NW configuration/UE request, including RS configuration for target CSI measurement and ID configuration for NW-side additional conditions
- Proposal 61a (Sec 9.1.4.1): Confirm necessity of NW-side monitoring based on target CSI via eType2 codebook contingent on UE capability, plus at least one additional monitoring option for UEs not supporting concurrent eType2 and CSI compression
- Proposal 64b (Sec 9.1.4.1): For NW-first training in options 3, 4, and 5, conclude root cause identification achievable through ground-truth CSI reporting, with NW comparing inference using NW vs UE side CSI generation parts
- Proposal 1 (Huawei - Sec 9.1.4.1): For Direction A Option 3a-1/4-1, study performance target including end-to-end, encoder-only, and decoder-only metrics, with testing dataset from NW for performance target verification