R1-2410021
discussion
On AI/ML for CSI compression
From Lenovo
Summary
Lenovo's technical contribution on AI/ML for CSI compression addresses data collection, model monitoring, inter-vendor collaboration, and quantization schemes. The document contains 19 specific proposals across 6 major technical areas, with emphasis on two-sided model training and deployment challenges.
Position
Lenovo advocates FOR Direction A inter-vendor collaboration (UE-side offline engineering) over Direction B (direct parameter use), strongly pushing AGAINST Direction C (fully standardized models) due to performance limitations and specification complexity. They prioritize practical deployment solutions with local decoder construction first, comprehensive model monitoring with root-cause analysis, and hybrid quantization schemes combining scalar and vector quantizers for optimal performance.
Key proposals
- Proposal 1 (Sec 2.1): Support procedures/signaling enabling UE/NW to associate the data/samples with the conditions/additional conditions under which the data/samples has been collected
- Proposal 2 (Sec 2.2): Support procedures/signaling enabling UE/NW for transmission of subset of samples among the set of measured/collected samples from the environment
- Proposal 3 (Sec 2.3): For transmission of ground-truth CSI samples, consider the performance of transmitting more samples, instead of fewer samples with higher resolution per sample
- Proposal 5 (Sec 3.1): Study mechanism to determine the main contributor(s) of the lower performance of the model at least for issues related to deployed encoder/decoder model, communication link, or data-drift
- Proposal 6 (Sec 3.1.1): Study mechanism for root-cause determination based on exchange of some test data-set between the NW and the UE
- Proposal 10 (Sec 4.1.1): Due to performance limitation and also the required high specification effort, we suggest deprioritizing Direction C for inter-vendor training collaboration
- Proposal 11 (Sec 4.1.2): For option 4-1 and 3a-1 of Direction A, prioritize schemes based on first construction of the 'local' decoder and then training of the encode model
- Proposal 15 (Sec 4.1.3): Until further investigation, give higher priority to options based on Direction A over options based on Direction B
- Proposal 16 (Sec 4.2): Support definition of pairing information based on the conditions/additional conditions assigned to the samples of the datasets used for training of the model
- Proposal 17 (Sec 4.3): Further study model identification/selection procedures during inference time when different models have been developed for different UE-NW vendor pairs
- Proposal 18 (Sec 5): Prioritize Case 2 and Case 3 for temporal domain aspects of AI/ML-based CSI compression using two-sided model
- Proposal 19 (Sec 6): Support procedures/signalling enabling CSI-compression models having both Scaler and vector Quantizers for generation of the CSI-feedback bits