R1-2409744
discussion
Other aspects of AI/ML model and data
From Intel
Summary
Intel presents 15 proposals and 7 observations regarding AI/ML model identification and transfer mechanisms for NR Rel-19, specifically focusing on CSI compression use cases. The document argues for model-ID-based identification to enable granular Life Cycle Management (LCM) and proposes prioritizing model transfer/delivery Case z4 (open format, known structure) while deprioritizing Case z1 (proprietary format). It details procedures for associating model IDs with data collection configurations and datasets to facilitate inter-vendor collaboration.
Position
Intel proposes supporting model-ID-based identification for two-sided CSI compression models to enable finer granularity in Life Cycle Management (LCM) compared to functionality-level identification. They argue that MI-Option 1 (data collection configuration), MI-Option 2 (dataset transfer), and MI-Option 3 (model transfer from NW) are all applicable and beneficial for CSI compression, with specific emphasis on UE-specific model IDs for dataset-based identification. Intel presents a technical case against model transfer/delivery Case z1, proposing its deprioritization in Rel-19 due to the burden of offline cross-vendor collaboration and limited benefits over Case y. Conversely, they propose supporting Case z4 (open format, known structure) and Case y, suggesting the specification of a family of known model structures to alleviate inter-vendor alignment burdens. They further propose that for Case z4, the network should transfer parameters for known structures reported by the UE, with support for partial parameter updates and a defined minimum application time for inference readiness.
Key proposals
- Proposal 1 (Functionality and Model Identification): Consider support of model-ID-based identification for two-sided models for CSI compression by enabling the network to provide a model ID to the UE for model identification type B.
- Proposal 2 (On MI-Option 1): For model identification with data collection configurations, consider options where model IDs are either determined prior to the assignment of an Associated ID or assigned at the time of association.
- Proposal 3 (On MI-Option 1): MI-Option 1 is applicable and beneficial for CSI compression using two-sided models, enabling model-level granularity for LCM operations.
- Proposal 4 (On MI-Option 2): For model identification with dataset transfer, model IDs should be UE-specific and can be assigned/reported by the UE or determined based on a relationship with a provided 'ID-X' for the dataset.
- Proposal 5 (On MI-Option 2): To alleviate inter-vendor coordination reliance, specify candidate values for dataset characteristics such as format, size, and normalization details.
- Proposal 6 (On MI-Option 2): MI-Option 2 is applicable for two-sided CSI compression models, including generalized and localized models trained at the UE-side or UE-side OTT server.
- Proposal 7 (On MI-Option 3): For model identification in model transfer from NW to UE, the UE performs identification to request a model and ID, with transfer provisioned in response to explicit or implicit requests.
- Proposal 8 (Model Transfer/Delivery): Deprioritize model transfer/delivery Case z1 in Rel-19 due to limited benefit compared to Case y, large offline cross-vendor collaboration burden, and storage burdens on the 3GPP network.
- Proposal 9 (Model Transfer/Delivery): Support model transfer/delivery Case y and Case z4 in Rel-19, considering specifying a group/family of model structures/backbones for Case z4 to reduce inter-vendor alignment burdens.
- Proposal 10 (Model Transfer/Delivery): For Case z4, follow procedure Alt. A where the UE reports supported known model structures and the NW transfers parameters for those structures.
- Proposal 11 (Model Transfer/Delivery): Assume models based on transferred parameters are ready for inference after a specified minimum application time.
- Proposal 12 (Model Transfer/Delivery): Use a 'first indication' to identify a known model structure and a 'second indication' to identify the model ID for parameters transferred for that structure.
- Proposal 13 (Model Transfer/Delivery): Support Case z4 involving the transfer of only a part of the parameters for a model of known structure, allowing for incremental updates.
- Proposal 14 (Model Transfer/Delivery): Align 'known model structures' for Case z4 via specified candidates, inter-vendor collaboration, or a combination of both.
- Proposal 15 (Model Transfer/Delivery): Specify candidate values for model input/output related information for CSI compression, including quantization method, encoder-decoder interface, fixed point representation, and compression type.