R1-2407799
discussion
Discussion on study for AI/ML CSI compression
From ZTE
Summary
ZTE provides comprehensive analysis on AI/ML CSI compression inter-vendor collaboration approaches, presenting 27 proposals across three main directions (UE-side offline engineering, on-device operation, fully standardized models) and addressing remaining specification issues including data collection, CQI determination, and performance monitoring.
Position
ZTE advocates for prioritizing Option 3 (standardized reference model structure + parameter exchange) with NW-first training and over-the-air delivery as the most feasible inter-vendor collaboration approach. They push against Option 4 due to huge dataset exchange overhead and UE-side monitoring due to complexity concerns. ZTE strongly supports down-selecting between Case 2 and Case 3 to reduce specification efforts and favors enhanced Rel-16 eTypeII codebook for data collection over more complex alternatives.
Key proposals
- Proposal 1 (Sec 2): To reduce the efforts for specification impact analysis and standardization, conduct comparison between Case 2 and Case 3 for potential down selection.
- Proposal 4 (Sec 3.1): For Option 3a/5a of direction A, performance target and additional training dataset comprising target CSI should be shared from NW-side to UE-side to enable UE-side encoder training, validation, and testing.
- Proposal 8 (Sec 3.2): For direction B, the number of parameters associated with the parameter exchange is estimated to range between 1 million to 13 million, which is acceptable for transmission via over-the-air interface signaling.
- Proposal 12 (Sec 3.3): For direction C, synthetic data generated under 3GPP's statistical channel model can be a starting point for reference model training.
- Proposal 16 (Sec 4.1): For Option 1, prioritize to study the case of standardized CSI generation part.
- Proposal 18 (Sec 4.5): For Option 3/4/5, prioritize to study NW-first training scheme and exchanging the parameters/model/dataset from NW side to UE side.
- Proposal 21 (Sec 5): For comprehensive analysis on AI/ML framework for two-sided model, further study a complete and unified solution for model identification, multi-vendor collaboration, and model pairing.
- Proposal 22 (Sec 6.1): For network side data collection, support to further study Enhanced Rel-16 eTypeII codebook design to achieve high-resolution CSI for model training and performance monitoring.
- Proposal 24 (Sec 6.2): For CQI determination, at least prioritize the specification impact discussions on Option 1a, Option 1b.
- Proposal 25 (Sec 6.3): Prioritize to study the specification impacts on at least the following case for model performance monitoring, NW-side monitoring based on the target CSI with realistic channel estimation associated to the CSI report, reported by the UE.
- Proposal 26 (Sec 6.3): In CSI compression using two-sided model use case, deprioritize the study on UE-side monitoring in Rel-19 study phase.
- Proposal 9 (Sec 3.2): For direction B, it is infeasible to implement a universal encoder across various UEs due to the inherent diversity in model structure preferences, data pre-processing requirements, hardware capabilities, power constraints, and quantization methodologies among different UE or chipset vendors.
- Proposal 17 (Sec 4.2): For Option 3, further study and evaluate the actual end-to-end performance of different sub-options when either side and/or both sides perform offline secondary development before actual system use.
- Proposal 19 (Sec 4.5): For Option 3/4/5, prioritize to study over-the-air delivery scheme.
- Proposal 27 (Sec 6.3): In CSI compression using two-sided model use case, further study a set of appropriate monitoring methodology and reasonable monitoring metrics for the given KPI under rank>1.