R1-2410719
discussion
Summary#1 of Additional study on AI/ML for NR air interface: CSI compression
From Qualcomm
Summary
This technical document from Qualcomm serves as a moderator summary for additional study on AI/ML-based CSI compression for NR air interface, containing over 120 proposals from various companies across temporal domain aspects, localized models, inter-vendor training collaboration, and monitoring approaches. The document presents comprehensive discussions and company positions on extending spatial/frequency compression to temporal domains while addressing inter-vendor collaboration challenges.
Position
Qualcomm, as the moderator, advocates FOR prioritizing temporal domain Cases 2 and 3 with separate prediction and compression as baseline, supporting Direction A (dataset sharing via option 4-1) combined with Direction C for inter-vendor collaboration, and emphasizing Transformer-based model structures. They push AGAINST Direction B (direct parameter sharing) due to feasibility concerns and advocate for deprioritizing complex inference aspects like CQI determination that don't affect core feasibility.
Key proposals
- Proposal 8 (Sec 1.1): For the additional potential spec impact of temporal domain CSI compression Case 2, consider methods to handle the misalignment of the accumulated CSI between NW part model and UE part model due to UCI missing
- Proposal 21a (Sec 1): 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 1): For temporal domain aspects Case 3, take separate prediction and compression (SPC) as baseline, study its LCM aspects and specification impacts. FFS: Joint prediction and compression
- Proposal 9 (Sec 2): For temporal domain CSI compression Case 3, potential spec impact may be needed for data collection of the CSI compression model and monitoring of the CSI prediction model
- Proposal 7 (Sec 3): In case of N different local regions with N different localized models, average performance should be considered over the N local regions
- Proposal 41a (Sec 4): Send LS to RAN2 for study of feasibility of parameter/dataset exchange using over-the-air and other signaling approaches
- Proposal 44a (Sec 4): Deprioritize direction B in Rel-19 study item considering no consensus on feasibility of developing UE-common or UE-specific encoder
- Proposal 46a (Sec 4): Prioritize Direction A and Direction C. Further study direction C to ensure minimum performance with multiple FFS aspects
- Proposal 47b (Sec 4): For studying standardized model structure, adopt Transformer as backbone for Case 0 spatial-frequency domain input, with specific adaptations for Cases 2 and 3
- Proposal 51b (Sec 5): Confirm necessity of ground-truth reporting for NW-side data collection using Rel-16 eType2 or Rel-18 eType2, FFS if enhancement needed
- Proposal 61a (Sec 6): Confirm necessity of NW-side monitoring based on target CSI reported via legacy eT2 codebook, plus additional monitoring option for non-supporting UEs
- Proposal 71a (Sec 7): Deprioritize discussion of CQI determination, quantization alignment, and CSI principles as not critical for determining feasibility
- Proposal 81a (Sec 8): Study performance-complexity trade-off comparing different AI/ML models in spatial-frequency vs angle-delay domains with FLOP and latency considerations