R1-2409784
discussion
Additional study on other aspects of AI model and data
From NVIDIA
Summary
NVIDIA argues for the necessity of deterministic, physics-based propagation modeling (specifically ray tracing) for accurate AI/ML data generation in 5G-Advanced and 6G. The document proposes concluding the need for model identification in Lifecycle Management (LCM) for two-sided models, collecting UE-sided training data, and standardizing model transfer solutions, specifically continuing studies on Cases y, z1, and z4.
Position
NVIDIA argues that deterministic, physics-based modeling, specifically ray tracing, is indispensable for generating accurate training data, presenting a technical case against relying solely on stochastic channel models which lead to poor test accuracy in site-specific deployments. They propose concluding the need for model identification in the context of Lifecycle Management (LCM) for two-sided models and explicitly request studying MI-Option 3, MI-Option 4 (for CSI compression reference models), and MI-Option 5. NVIDIA supports the continuation of studies for model transfer Cases y, z1, and z4, while acknowledging the deprioritization of Cases z2, z3, and z5. They align with agreements to study specific procedures for MI-Option 1 (AI-Example1) and MI-Option 2 (AI-Example2-1), including the use of associated IDs for data collection consistency within a cell. Furthermore, they support defining 'known model structure' in Case z4 to include specific neural network parameters such as layer types, sizes, and connections.
Key proposals
- Observation 1 (Sec 2): Deterministic, physics-based modelling for wireless propagation, especially ray tracing, are essential for studying, evaluating, and developing AI/ML models in 5G-Advanced toward 6G.
- Proposal 1 (Sec 3): Conclude that there is a need for model identification in the context of LCM for two-sided models.
- Proposal 2 (Sec 3): Besides MI-Option 1 and MI-Option 2, describe examples for MI-Option 3 (model transfer), MI-Option 4 (standardization of reference models for CSI compression), and MI-Option 5 (model monitoring) to study their feasibility/necessity.
- Proposal 3 (Sec 4): Conclude that there is a need for collection of UE-sided model training data.
- Proposal 4 (Sec 5): Conclude that there is a need to consider standardised solutions for transferring/delivering AI/ML models.
- Proposal 5 (Sec 5): Continue to study Cases y, z1 and z4 for transferring/delivering AI/ML models.
- Agreement (Appendix A.2): Study AI-Example1 for MI-Option 1, where NW signals data collection configurations with associated IDs, UE collects data, develops models, and reports model information/IDs to NW.
- Agreement (Appendix A.3): For Case z4, study Alt A (UE reports supported known model structures, NW transfers parameters) and Alt B (NW indicates candidate structures, UE reports support, NW transfers parameters).
- Agreement (Appendix A.3): For MI-Option 2 (two-sided model), study AI-Example2-1 where NW transfers a dataset to UE, UE develops model part, and UE reports model information to NW.
- Conclusion (Appendix A.4): Model identification is at least applicable to some inter-vendor training collaboration options of CSI compression using two-sided models.
- Conclusion (Appendix A.4): The model identification procedure dedicated to MI-Option 5 is not pursued for Rel-19 normative work.
- Conclusion (Appendix A.4): The model identification procedure dedicated to MI-Option 2 for one-sided model is not pursued for Rel-19 normative work.
- Agreement (Appendix A.4): Confirm Working Assumption that UE assumes NW-side additional conditions with the same associated ID are consistent at least within a cell.
- Conclusion (Appendix A.5): Model transfer/delivery Case z2 is deprioritized for two-sided model in Rel-19.
- Agreement (Appendix A.5): For Case z4 inference, study whether/when UE needs transfer of new parameters, when model is ready for inference, and transfer of partial parameters.