R1-2410915
LS out
[Draft] LS on signalling feasibility of dataset and parameter sharing
From Qualcomm
Summary
This liaison statement from Qualcomm/RAN1 to RAN2 requests feedback on standardized signaling feasibility for AI/ML-based CSI compression inter-vendor collaboration, specifically for sharing datasets and model parameters between network-side and UE-side entities. The document contains multiple company comments but no formal proposals, focusing on three main options for data/parameter exchange with sizes ranging from 37.5MB to 450MB.
Position
Qualcomm/RAN1 is advocating FOR standardized signaling mechanisms to enable inter-vendor collaboration in AI/ML-based CSI compression through dataset and parameter sharing between network-side and UE-side entities. They are pushing FOR feasibility assessment of three specific options (4-1, 3a-1 with/without target CSI) with relaxed latency requirements (days/weeks) and infrequent sharing. They position this as essential to resolve inter-vendor training collaboration complexity while maintaining that proprietary information concerns are minimal for the proposed options.
Key proposals
- Request to RAN2: Feedback on standardized signaling feasibility (over-the-air and/or other approaches) for dataset sharing consisting of {Target CSI, CSI feedback}
- Request to RAN2: Feedback on standardized signaling feasibility for encoder parameter sharing
- Request to RAN2: Feedback on standardized signaling feasibility for encoder parameter sharing + dataset sharing consisting of {target CSI}
- Option 4-1: Dataset sharing containing N1 samples of {target CSI, CSI feedback} corresponding to input and output of nominal encoder
- Option 3a-1 without target CSI: Sharing parameters corresponding to nominal encoder
- Option 3a-1 with target CSI: Sharing parameters corresponding to nominal encoder, along with dataset containing N2 samples of {target CSI}
- OPPO comment: Suggest indicating updated dataset sharing condition where only new updated datasets are exchanged instead of whole datasets
- ETRI comment: Consider quantization/compression for model parameters using Float16 or Int8 to reduce model size
- Ericsson comment: SA2 and SA5 should be in 'To' list rather than 'CC' list for non-over-the-air signaling feasibility
- Huawei comment: Focus on NW-side to UE-side server parameter/dataset sharing and assume dataset range from 100K to 600K samples
- MediaTek comment: Use float32 instead of float16 for calculations as it's the default in PyTorch and TensorFlow
Revision chain
2 versions in this meeting · oldest first
-
R1-2410915 ← you are here LS out revised
-
LS on signalling feasibility of dataset and parameter sharing