R1-2409783
discussion
Additional study on AI-enabled CSI compression
From NVIDIA
Summary
NVIDIA argues that stochastic channel models are insufficient for demonstrating AI/ML CSI compression gains, proposing instead the use of site-specific models trained on data generated via ray tracing in defined reference scenarios. The document outlines six proposals covering performance study methodologies, inter-vendor training collaboration feasibility, and post-deployment performance monitoring mechanisms.
Position
NVIDIA argues that stochastic channel models are insufficient for demonstrating AI/ML CSI compression gains, proposing the consideration of site-specific AI/ML models. They propose defining a common reference scenario with site specificity, selecting between real-scenario maps or synthetic-scenario maps, and using ray tracing to generate channel data for model development. Regarding inter-vendor training, NVIDIA concludes that it is feasible to resolve collaboration issues, highlighting Training Type 3 (separate training) as a pragmatic option to safeguard proprietary information. They propose studying post-deployment performance monitoring mechanisms to detect performance degradation and non-compliance. NVIDIA presents multiple options for inter-vendor collaboration, including fully standardized reference models, standardized datasets, and parameter/model exchanges.
Key proposals
- Proposal 1 (Sec 3): Site-specific AI/ML models for CSI compression should be considered to improve performance gain.
- Proposal 2 (Sec 3): Define a common reference scenario with site specificity as a basis for further study of AI/ML based CSI compression.
- Proposal 3 (Sec 3): Select one of two options to define a common reference scenario: Option 1 (Real-scenario map) or Option 2 (Synthetic-scenario map).
- Proposal 4 (Sec 3): With a common reference scenario with site specificity, ray tracing is used to generate channel data for the development and evaluation of site-specific AI/ML models for CSI compression.
- Proposal 5 (Sec 4): RAN1 to conclude that it is feasible to resolve issues related to inter-vendor training collaboration for AI/ML-based CSI compression.
- Proposal 6 (Sec 5): RAN1 to study post-deployment performance monitoring mechanisms to detect performance degradation and non-compliance to guarantee satisfactory performance of AI/ML-based CSI compression in the field.
- Option 1 (Sec 4): Fully standardized reference model (structure + parameters) to eliminate inter-vendor collaboration complexity.
- Option 3 (Sec 4): Standardized reference model structure + Parameter exchange between NW-side and UE-side.
- Option 4 (Sec 4): Standardized data / dataset format + Dataset exchange between NW-side and UE-side.
- Option 5 (Sec 4): Standardized model format + Reference model exchange between NW-side and UE-side.
- Training Type 3 (Sec 4): Separate training at network side and UE side is identified as a pragmatic approach to safeguard proprietary information and eliminate collaboration during training iterations.
- Observation 1 (Sec 3/6): The Rel-18 study shows that using data generated from stochastic channel models is not sufficient to demonstrate the performance gain of CSI compression compared to non-AI/ML based algorithms.
- Option 1 (Sec 3): Real-scenario map that is a virtual representation of a real area on earth for defining common reference scenarios.
- Option 2 (Sec 3): Synthetic-scenario map that is artificially constructed to mimic a certain environment such as urban macro, rural macro, indoor office, or indoor factory.