R1-2410508
discussion
Additional study on AI/ML for NR air interface - CSI compression
From MediaTek
Summary
MediaTek's contribution addresses additional study aspects for AI/ML-based CSI compression in NR Release 19, presenting 10 technical proposals covering temporal-domain compression, error tolerance, inter-vendor collaboration approaches, model monitoring techniques, and data collection methods for training AI/ML models.
Position
MediaTek advocates FOR: (1) using temporal-domain CSI compression with separate AI/ML models for prediction and compression rather than unified models, (2) prioritizing AI/ML model transfer over dataset transfer for inter-vendor collaboration, (3) leveraging uplink CSI from SRS for network-side training instead of requiring downlink CSI feedback, (4) fine-grained associate IDs to enable cell-specific localized AI/ML models, and (5) Power Spectral Entropy (PSE) based monitoring techniques. They push AGAINST: unified prediction+compression AI/ML models that increase complexity, coarse-grained associate IDs that prevent localized optimization, and over-reliance on downlink CSI collection for training.
Key proposals
- Proposal 1 (Sec 2): Clarify the boundary of AI/ML model, i.e., in which step it starts or ends
- Proposal 2 (Sec 3): Evaluate the feedback error tolerance of eType II and compare it with that of AI/ML model
- Proposal 3 (Sec 4.1): For direction A, prioritize transferring AI/ML model rather than transferring dataset
- Proposal 4 (Sec 4.1): For direction A, prioritize transferring AI/ML encoder through 3a-1, either with or without {Target CSI}
- Proposal 5 (Sec 4.2): For Direction B, discuss how to consider UE's computational constraints when encoder is provided by NW
- Proposal 6 (Sec 4.2): For Direction B, discuss whether re-engineering on reference AI/ML models are allowed right before their deployment
- Proposal 7 (Sec 4.2): For direction B, limit the updates occurrence to specific occasions including UE-side detection and NW-side decision based triggers
- Proposal 8 (Sec 5.1.2): Discuss NW-side AI/ML model I/O-based monitoring using uplink CSI samples collected from SRS
- Proposal 9 (Sec 5.2): Deprioritize coarse granularity of associate ID, i.e., assigning one associate ID to multiple cells
- Proposal 10 (Sec 6): For NW-side AI/ML model training, NW can rely on UL CSI samples collected from SRS sent by UEs. Use of DL CSI can be limited to finetuning purposes