R1-2509148
discussion
AI/ML in 6GR Air Interface
From MediaTek
Summary
MediaTek presents 10 proposals and 8 observations for 6G AI/ML physical layer study, covering a unified modular AI framework, CSI compression/feedback (JSCCM), CSI prediction across spatial/frequency/time domains, UL DMRS overhead reduction with AI receivers, and handling of 5G NR use cases.
Position
MediaTek proposes a unified modular AI/ML framework for 6G air interface with a shared backbone model and multiple use-case-specific head models, demonstrating via feasibility study on CSI compression and positioning that this approach reduces UE-side complexity by up to 50% while maintaining or improving performance. They present technical case for JSCCM (Joint Source-Channel Coding and Modulation) over SSCC and JSCC for CSI feedback, showing JSCCM outperforms at all SNRs, and require priority for approaches limiting UE complexity such as a trainable linear transformation. For spatial CSI prediction, they observe that shifted and random CSI-RS patterns outperform uniform patterns and propose studying joint CSI-RS design and AI-based channel inference. They require deprioritizing AI-receiver at the UE-side for DL due to high inference complexity (128 M–10,856 M FLOPs in evaluated configurations) and restrict AI-receiver study to NW-side for UL transmission only. They also propose studying temporal-domain CSI prediction with different time resolution at prediction and observation window, and take Rel-19 AI/ML positioning sub-cases 1, 3a, 3b as feasible without further study while leaving open the possibility of updates to NR AI/ML use cases due to underlying 6GR design changes.
Key proposals
- Proposal 1 (Sec: AI/ML Framework for 6GR Air Interface): Study a general/unified AI/ML framework for 6GR Air Interface considering extensibility for emerging applications, leveraging synergy between use cases, facilitating both offline and online model (re-)training, avoiding unreasonable signalling overhead, and facilitating separate/partial training of AI/ML models.
- Proposal 2 (Sec: Modular unified architecture for similar use cases for efficient LCM): Consider exploiting similarities among use cases by utilizing a modular architecture across similar use cases and exploiting the knowledge of the modular structure to enable training, data collection and monitoring on a modular basis, instead of performing these operations separately for each use case.
- Proposal 3 (Sec: CSI Compression and Feedback Based on Joint Source-Channel Coding (JSCC) or Joint Source-Channel Coding and Modulation (JSCCM)): JSCCM offers performance advantages over both JSCC and SSCC with link adaptation in terms of SGCS and may be studied further in detail to evaluate its impact on the uplink transmit chain and its performance under realistic assumptions. Priority must be given to approaches which limit the complexity at the UE side, e.g., using a trainable linear transformation.
- Proposal 4 (Sec: Study of CSI Compression Based on Unified Architecture): Study the feasibility of AI/ML-based unified downlink CSI feedback framework, including but not limited to, its performance, training and inference complexity, inference latency, model generalization issues.
- Proposal 5 (Sec: CSI Prediction in Spatial and/or Frequency Domain): Study the feasibility of AI/ML-based channel inference/prediction for CSI-RS overhead reduction in spatial and/or frequency domains including the following sub-use cases and related details.
- Proposal 6 (Sec: CSI Prediction in Time Domain): Study temporal-domain AI/ML CSI prediction with different time resolution at prediction and observation window to reduce reference signaling overhead.
- Proposal 7 (Sec: Uplink DMRS Overhead Reduction with AI-Transceiver/Receiver): Deprioritize study of AI-receiver at the UE-side for DL reception due to the high inference complexity which would translate to high implementation complexity at the UE. Any 6G study related to AI-receivers should be restricted to inference at the NW-side for UL transmission only.
- Proposal 8 (Sec: Generalization Study for DMRS-free use case): For AI-receiver-related use cases, detailed generalization studies should be carried out to ascertain their robustness under different channel conditions/scenarios.
- Proposal 9 (Sec: Handling of 5G NR Use Cases): While NR use cases are not intended to be evaluated further for the purpose of understanding their feasibility, it does not preclude underlying 6GR design changes compared to NR and it is understood that such underlying enhancements may warrant updates to the past AI/ML modelling/evaluations/assumptions for the NR use cases at a later stage of the 6G AI/ML study.
- Proposal 10 (Sec: Handling of 5G NR Use Cases): Take the sub-cases 1, 3a and 3b in Rel-19 AI/ML positioning as a feasible use case without the need for further study.