R1-2409450
discussion
AI/ML for CSI compression
From Ericsson
Summary
Ericsson presents a comprehensive analysis of inter-vendor training collaboration options for AI/ML-based CSI compression, arguing for the use of 3GPP synthetic data and standardized phase normalization to ensure interoperability. The document evaluates three main directions (A, B, and C), concludes that UE-side first training is unsupported, and emphasizes NW-side performance monitoring using high-resolution target CSI reporting while highlighting significant complexity concerns for temporal domain compression cases.
Position
Ericsson requires that reference models for inter-vendor collaboration (Options 1, 3a, 3b) be designed using 3GPP channel model based synthetic data rather than field data, citing excessive work and bias risks. They oppose UE-side first training for Options 3/4/5, arguing it necessitates multiple parallel models at the NW-side, and they reject over-the-air delivery for Direction A due to signaling overhead and latency. To address dataset distribution mismatch, Ericsson proposes standardizing the phase normalization of precoding vectors and using mixed datasets for training. For performance monitoring, they conclude that UE reporting of high-resolution target CSI is necessary to enable NW-side intermediate KPI monitoring and error cause detection, specifically proposing enhancements to the eType-II format. Finally, they highlight that AI CSI compression Case 3 incurs computational complexity 200-300 times higher than Rel-18 eType II, necessitating further study on complexity reduction.
Key proposals
- Proposal 1 (Sec 2.1): RAN1 conclude that reference models for inter-vendor collaboration options 1, 3a, and 3b are designed and specified using 3GPP channel model based synthetic data.
- Proposal 2 (Sec 2.2): For Direction B (Option 3b), model parameter quantization and input data pre-processing must be standardized together with the CSI generation model structure to mitigate inter-vendor collaboration complexity.
- Proposal 4 (Sec 2.3): UE-side first training cases for RAN1 Option 3/4/5 are not supported as they necessitate multiple models to operate in parallel at the NW-side.
- Proposal 5 (Sec 2.3.1): For Option 3a-1, in addition to CSI generation model parameters, {Target CSI} as training/testing dataset, performance requirements, and pairing information are necessary to be shared from NW-side to UE-side.
- Proposal 8 (Sec 2.3.3): For Direction A, the over-the-air delivery method for exchanging information from the NW-side to UE-side is not supported due to high complexity and overhead.
- Proposal 10 (Sec 2.4.2): To solve data distribution mismatch issues, standardize the phase normalization of a precoding vector used as input to an encoder for CSI compression using two-sided model use case.
- Proposal 12 (Sec 2.4.3): Study how to represent a standardized reference model (structure + parameters) and/or a standardized reference model structure in 3GPP specifications.
- Proposal 13 (Sec 3.1): Conclude that it is necessary to specify UE reporting high resolution target CSI to enable NW-side intermediate KPIs based performance monitoring and performance degradation error cause detection.
- Proposal 14 (Sec 3.1): Capture in TR that ground-truth CSI report based on enhancements of the eType-II format with new parameters shall be defined to ensure high-accuracy model performance monitoring.
- Proposal 16 (Sec 4.1): Further study methods to reduce the AI computational complexity for AI CSI compression Case 3, given the high FLOP ratio compared to legacy schemes.