R1-2409671
discussion
Discussion on CSI compression
From vivo
Summary
This document from vivo analyzes inter-vendor training collaboration for AI/ML-based CSI compression in NR, covering Directions A (UE-side offline engineering), B (NW-side encoder sharing), and C (standardized reference models). It presents 25 proposals and 17 observations, arguing that Direction A requires careful handling of data distribution mismatches and proprietary concerns, Direction B is feasible with common encoders, and Direction C should prioritize standardizing the encoder structure using CNN or Transformer backbones.
Position
vivo argues that for Direction A, exchanging target CSI from NW to UE is unnecessary for option 3a-1 if the NW includes mixed UE data in training, but proprietary concerns regarding decoder structure disclosure for Alt 1 training must be studied. They propose addressing data distribution mismatch in Direction A through NW-side data drift detection and end-to-end model updates rather than relying solely on UE-side offline engineering. For Direction B, vivo asserts that a common encoder across UEs is feasible despite UE-side condition variations, and there are no significant overhead or proprietary concerns. Regarding Direction C, vivo proposes that the reference model must include at least the encoder part and recommends standardizing model structures starting with SF compression (Case 0) using CNN or Transformer backbones, leveraging RAN4 agreements as a starting point for hyperparameter alignment.
Key proposals
- Proposal 1 (Sec 2.1.1): For Direction A option 3a-1, additional information shared from NW to UE should include performance requirements, assistant information (e.g., model ID), and information related to collecting UE-side training datasets.
- Proposal 2 (Sec 2.1.2): For Direction A option 4-1, additional information may include performance requirements, assistant information, dataset collection info, model backbone/structure information, and an additional dataset for performance testing.
- Proposal 4 (Sec 2.1.3): RAN1 should further study proprietary information disclosure risks in Direction A, specifically regarding decoder structure/loss function disclosure for Alt 1 training and dataset sharing between UEs of different vendors.
- Proposal 5 (Sec 2.1.4): RAN1 should further study overhead in Direction A based on encoder size, number of encoders, target CSI size, and transfer frequency for option 3a-1, and dataset size/samples/transfer frequency for option 4-1.
- Proposal 7 (Sec 2.1.5): RAN1 may consider studying the necessity and feasibility of specifying some UE-side additional conditions (e.g., SVD phase normalization) to facilitate inter-vendor training collaborations.
- Proposal 8 (Sec 2.1.5): For Direction A, data distribution mismatch should be addressed via an NW-side approach involving data drift detection, NW-side data collection, end-to-end model updating, and subsequent UE-side model updating based on shared information.
- Proposal 9 (Sec 2.2.1): For Direction B, there is no UE-side overhead concern for transferring encoder parameters (estimated hundreds of Kbytes), and signaling design can be optimized to reduce overhead for large numbers of UEs.
- Proposal 10 (Sec 2.2.2): A common encoder for different UEs is feasible for Direction B, even with different UE-side conditions such as SVD phase normalization, antenna virtualization, and antenna spacing/layout.
- Proposal 11 (Sec 2.2.3): For Direction B, performance degradation causes can be identified via NW-side monitoring by checking SGCS of the original model, comparing expected vs. reported CSI feedback, and verifying NW-side decoder deployment.
- Proposal 13 (Sec 2.3.1): To achieve better field performance in Direction C, the reference model should at least include the encoder part.
- Proposal 14 (Sec 2.3.2): RAN1 should choose between Option 1 (RAN1 reference model is RAN4 reference model, requiring RAN4 to consider TSF use cases and scalability) or Option 2 (RAN1 reference model is distinct, with RAN1 handling deployment aspects like TSF and scalability).
- Proposal 17 (Sec 2.3.4): Standardization of model structure for CSI compression should start with TSF Case 0 (SF compression), as TSF Case 2 and 3 structures are closely related and can be generated based on Case 0.
- Proposal 18 (Sec 2.3.4): For studying standardization of encoder model structure in RAN1, Transformer and CNN should be considered as candidate backbones for Case 0.
- Proposal 23 (Sec 2.3.4): RAN1 can consider the CNN structure presented in RAN4 agreements as a starting point for discussing specified model structure.
- Proposal 25 (Sec 2.3.4): RAN1 should first focus on discussing model structure based on a set of parameters for evaluations, and then consider scalability over Tx ports, ranks, payload sizes, and bandwidths.