R1-2409880
discussion
Views on AI/ML model based CSI compression
From Xiaomi
Summary
Xiaomi presents comprehensive views on AI/ML-based CSI compression feedback for NR, covering performance evaluation of temporal domain compression (Case 2), inter-vendor collaboration options, performance monitoring, and specification impacts. The document contains 19 proposals and 14 observations addressing key technical challenges in two-sided AI/ML model deployment.
Position
Xiaomi strongly advocates for AI-TSF (temporal-spatial-frequency) compression over traditional AI-SF compression, demonstrating significant performance gains (8.9-13.71% over eType II codebook). They push for practical inter-vendor collaboration solutions, favoring over-the-air parameter exchange (Option 3a-1) over dataset exchange due to overhead concerns, and promoting common encoders to address proprietary information disclosure. They oppose UE-side performance monitoring for root cause analysis due to accuracy and overhead concerns, instead advocating for NW-side monitoring approaches.
Key proposals
- Proposal 1 (Sec 4): To align the missed historical CSI information when UCI loss occurs, the latest output historic CSI of encoder and decoder before missing UCI could be respectively adopted as the input of encoder and decoder in the current inference instance.
- Proposal 2 (Sec 6.1): For exchanged target CSI, Rel-16 eType II codebook or high-resolution eType II codebook could be used to transfer dataset for addressing the overhead issue.
- Proposal 4 (Sec 6.2): For saving overhead of transferring model parameters, the parameters of neural partial network layers of encoder are transferred from NW-side to UE-side, while the parameters of remained neural network layers are standardized.
- Proposal 7 (Sec 6.3): Suggest the collected dataset based on 3GPP's statistical channel model for reference model(s) training.
- Proposal 10 (Sec 7): Considering affordable overhead and limited inter-vendor collaboration complexity, at least encoder parameters exchanged via over-the-air from NW-side to UE-side could be supported for Option 3a-1.
- Proposal 11 (Sec 7): In order to make the two-sided model be compatible, the pairing information for model pairing is assigned by gNB with parameter/dataset exchange for Option 3a-1//4-1.
- Proposal 12 (Sec 7): UE capability on supporting different inter-vendor training collaboration options is reported to NW side.
- Proposal 14 (Sec 7): Quantization approach of CSI feedback needs to specify for Option 3a-1/3b/4-1/1.
- Proposal 15 (Sec 9): Considering the accuracy of monitoring and signalling overhead, identifying the cause of performance degradation based on UE-side performance monitoring is studied with low priority.
- Proposal 16 (Sec 9): In order to identify the cause of the performance degradation, both target CSI and CSI feedback measured by UE are reported for NW-side performance monitoring.
- Proposal 17 (Sec 11.1): For NW side data collection, ground truth CSI is reported by using codebook, e.g., Rel-16 eType II codebook or eType II codebook with enhanced parameters; Cell-specific CSI-RS resource is configurated by NW-side to collect training dataset.
- Proposal 18 (Sec 11.1): For UE side data collection, NW configuration or UE request could be considered to collect dataset at UE side; The configuration of temporal aspects for Rel-18 Type II Doppler codebook could be as a starting point for Case 2/3.
- Proposal 19 (Sec 11.2): Support CQI calculated using two stage approach.
- Proposal 3 (Sec 6.1): UE first training a nominal decoder to address performance impact due to mismatch between NW side data distribution and UE side data distribution could be considered.
- Proposal 6 (Sec 6.2): A common encoder is used by users to address the UE's proprietary information concern of disclosing encoder parameters.