R1-2410218
discussion
Further views on consistency issues in CSI prediction
From Sony
Summary
Sony's contribution addresses consistency issues between AI/ML model training and inference for CSI prediction in NR systems, presenting 2 specific proposals for RAN1 study. The document identifies how differences in UE capabilities, network configurations, and model functionality configurations can cause performance degradation during inference.
Position
Sony advocates FOR comprehensive study of training-inference consistency issues in AI/ML CSI prediction models, pushing for systematic characterization and resolution of inconsistencies caused by UE capability differences, network configuration mismatches, and functionality configuration variations. They are advocating AGAINST proceeding with normative work without properly addressing these fundamental consistency challenges that can cause significant performance degradation.
Key proposals
- Proposal 1 (Sec: Consistency Issues): RAN1 should study whether differences in UE capabilities and/or semi-static network configurations can engender inconsistency issues between model training and inference and how these can be resolved
- Proposal 2 (Sec: Consistency Issues): RAN1 should study whether difference in functionality configuration of models during training and inference can lead to inconsistencies and how to resolve these