R1-2409731
discussion
Discussion on other aspects of AI/ML
From Ericsson
Summary
Ericsson analyzes model identification options for two-sided AI/ML models in NR, specifically focusing on CSI compression use cases. The document presents 7 proposals and 6 observations, arguing that over-the-air dataset delivery is infeasible and recommending that model identification discussions be postponed until the underlying CSI compression use case matures.
Position
Ericsson opposes the over-the-air delivery of datasets from the network to the UE for two-sided model training, citing high complexity, signaling overhead, and questionable feasibility. They propose that model identification discussions for MI-Options 2, 3, and 4 be postponed until the two-sided CSI compression use case resolves issues regarding data distribution mismatch and inter-vendor interoperability. For model identification, Ericsson proposes that model IDs be generated locally by the network, incorporating unique vendor, location, and site IDs alongside proprietary components, rather than using a central global registry. Regarding model transfer, they prioritize 'case y' (minimal specification impact) and only suggest considering 'case z4' (specified model structure and coefficient precision) if NW-sided training collaboration is deemed infeasible. They argue that without access to NW decoder outputs, end-to-end performance verification is challenging, and vendor-specific conformance testing would break 3GPP-level multi-vendor interoperability.
Key proposals
- Proposal 1 (Introduction/Conclusions): For model identification of two-sided models, RAN1 study focuses only on MI-options 2, 3, and 4.
- Proposal 2 (MI-Option 2): Over-the-air delivery method for exchanging dataset from the NW-side to UE-side is not supported; RAN1 should request relevant WGs to assess the feasibility of other standardized approaches for dataset exchange.
- Proposal 3 (MI-Option 2): For UE side part training based on dataset exchange, the model ID composes of a dataset ID generated locally by the NW, dependent on unique vendor/location/site IDs and a proprietary part.
- Proposal 4 (MI-Option 3): For UE side part training based on model exchange, the model ID is generated locally by the NW, dependent on unique vendor/location/site IDs and a proprietary part.
- Proposal 5 (MI-Option 4): For MI-Options 2, 3, and 4, RAN1 to conclude that there is no need to discuss until further progress is made for the two-sided CSI compression use case.
- Proposal 6 (Discussion on model transfer/delivery): Rel-19 RAN groups prioritize case y for model delivery if a need arises based on use case progress, and down-prioritize the other cases.
- Proposal 7 (Discussion on model transfer/delivery): Only if the collaboration burden of case y with NW-sided training is deemed infeasible, prioritize case z4 with specified model structure and coefficient precision.