R1-2509115
discussion
On AI/ML for 6G air interface
From Apple
Summary
This document from Apple presents 6 proposals and 1 observation across 5 sections addressing AI/ML for 6G air interface, covering use cases like CSI prediction for NW energy saving type 2, CSI compression with JSCCM, cross-frequency CSI prediction, AI/ML LCM framework enhancements, and next-step discussion prioritization.
Position
Apple proposes further study of CSI prediction across different antenna to port virtualization to enable type 2 spatial domain NW energy saving with low CSI-RS overhead, using eigen-vector SGCS and SGCS ratio with e-type 2 codebook as KPI. They propose adding SNR as additional model input for JSCM encoder/decoder training and inference, demonstrating up to 17% SGCS gain over e-type 2 in high-SNR regions. They require clarifying that e-type II codebook is used for SGCS calculation in cross-frequency CSI prediction and define the KPI as the SGCS ratio (SGCS_1/SGCS_2) to separate prediction loss from compression loss. For the LCM framework, they propose using 5G AI/ML LCM as baseline while requiring PLMN-unique Association ID/Pairing ID management and a simplified, scalable APU framework applicable beyond CSI reporting. They propose prioritizing day-one essential use cases for next-phase evaluation and maintaining a separate AI agenda for cross-use-case general framework discussions.
Key proposals
- Proposal 1 (Sec 2.1): Further study CSI prediction across different antenna to port virtualization to enable efficient, low overhead type 2 spatial domain NW energy saving.
- Observation 1 (Sec 2.2): Similar complexity/performance trade-off is observed in JSCM as the AI based CSI compression in NR.
- Proposal 2 (Sec 2.2): Update Table D sub-case B to add SNR as additional input information for model input for training and inference of the encoder/decoder.
- Proposal 3 (Sec 2.3): Update table B sub-case C to clarify that the e-type II codebook is used for SGCS calculation and define the KPI as the SGCS ratio = SGCS_1/SGCS_2.
- Proposal 4 (Sec 3): For 6G SI on AI/ML LCM framework, use 5G AI/ML LCM framework as starting point for one-sided and two-sided models, define management entity for PLMN unique Association ID/Pairing ID, and study a simplified scalable APU framework.
- Proposal 5 (Sec 4): Use cases that are part of essential procedures or provide the greatest benefits for day-one deployment can serve as the starting point for next phase discussions.
- Proposal 6 (Sec 4): Maintain a separate AI agenda to discuss the general framework, including at least the APU framework and association ID/pairing ID management.