R1-2409584
discussion
Views on additional study for AI/ML based CSI compression
From Samsung
Summary
Samsung presents views on further studies for AI/ML-based CSI compression in Rel-19, focusing on temporal aspects (Case 2 and Case 3), performance-complexity trade-offs, and inter-vendor training collaboration. The document contains 18 proposals and 16 observations, arguing that angle-delay (W2) domain compression offers superior generalization and robustness against data distribution mismatches compared to spatial-frequency domain compression.
Position
Samsung proposes that angle-delay (W2) domain compression is significantly more robust to data distribution mismatches than spatial-frequency (W) domain compression, citing up to 37.9% degradation for W-domain versus only 0.7% for W2-domain when mixing deployment scenarios. They require the study of an 'associated ID' mechanism to align UE-side and network-side training datasets with respect to network-side additional conditions, such as deployment scenarios or network-part model versions, for both Direction A and Direction C inter-vendor collaboration. Samsung concludes that data distribution mismatch due to UE antenna spacing has negligible impact on performance, allowing independent UE-side model updates. They propose using a relative KPI (SGCS_AI / SGCS_Baseline) for network-side monitoring to specifically identify AI/ML-related performance losses. Additionally, they propose studying Joint Source-Channel Coding and Modulation (JSCCM) to mitigate the cliff effect and reduce complexity.
Key proposals
- Proposal 1 (Sec 2.1): For Case 2 temporal aspects, consider two options for past CSI: past CSI generated by the AI/ML model or SD/FD basis vectors with angle-delay domain compression.
- Proposal 2 (Sec 2.1): For Case 2 with SD/FD basis as past CSI, consider reporting SD/FD basis per N CSI reporting occasions (N times longer periodicity) to reduce overhead.
- Proposal 4 (Sec 2.1): For Case 3 prediction, consider prediction instances in spatial-frequency-time, angle-delay-time, or angle-delay-Doppler domains.
- Proposal 5 (Sec 2.1): Model CSI dropping probabilities for Part II CSI groups (x%, y%, z%) to evaluate performance impact of error propagation in Case 2.
- Proposal 6 (Sec 2.1): Study the impact of input pre-processing (dimensionality reduction) on performance, model complexity, and generalizability.
- Proposal 7 (Sec 2.1): Study the impact of the number of SD/FD basis vectors on the performance-complexity trade-off in angle-delay domain compression.
- Proposal 8 (Sec 2.2): For ground-truth CSI collection in temporal cases, consider high-resolution codebook quantization including temporal aspects or explicit channel matrices for prediction cases.
- Proposal 9 (Sec 2.2): Study the necessity and specification impact of indicating network-side additional conditions for NW-first training with dataset exchange.
- Proposal 10 (Sec 2.2): Consider a KPI for performance monitoring defined as the ratio of SGCS for AI/ML-based CSI to baseline CSI, replacing the intermediate KPI in Rel-18 evaluation mechanisms.
- Proposal 13 (Sec 2.3): Conclude that data distribution mismatch causes considerable degradation for spatial-frequency domain inputs but negligible impact for angular-delay (W2) domain inputs in Direction C.
- Proposal 14 (Sec 2.3): Consider an associated ID to align UE-side and NW-side training dataset distributions with respect to NW-side additional conditions like deployment scenario.
- Proposal 15 (Sec 2.3): Conclude that data distribution mismatch resulting from UE antenna spacing does not impact performance, allowing independent UE-side model training.
- Proposal 16 (Sec 2.3): For Direction A, consider an associated ID to align UE-side and NW-side training datasets with respect to NW-side additional conditions.
- Proposal 18 (Sec 2.3): Consider CNN as a candidate backbone for the standardization of model structure, with FFS on applicability and quantization.