R1-2410380
discussion
Discussion on other aspects of AI/ML model and data
From NTT DOCOMO
Summary
NTT DOCOMO presents a comprehensive discussion on AI/ML model identification, data collection, and transfer mechanisms for 5G NR air interface, making 9 proposals and 6 observations covering scenario-specific models, model identification procedures, and standardization approaches.
Position
NTT DOCOMO strongly advocates FOR scenario/site-specific AI/ML models over generalized models, arguing they provide superior performance by learning specific environment tendencies that are difficult to mathematically model. They push FOR hybrid approaches combining multiple model identification options and standardized model transfer mechanisms. They are AGAINST deprioritizing model identification due to management complexity, and AGAINST case z1 model delivery unless clear gains are demonstrated over UE-side training approaches.
Key proposals
- Proposal 1 (Sec 2.1): 3GPP should consider the framework to support scenario/site specific model
- Proposal 2 (Sec 2.2): Future compatibility with model transfer and model storage at NW side should be taken into consideration after they are supported in 3GPP
- Proposal 3 (Sec 2.2): Since MI-Option1 does not help two-side model pairing, MI-Option1 should be used together with other MI-Options (e.g., MI-Option 2/3/4) for two-sided model
- Proposal 4 (Sec 4): Deprioritize case z1, unless explicit gain of case z1 compared to case y with UE side training is observed
- Proposal 5 (Sec 4): UE should report the indication that transferred model is ready for inference, when the compiling is necessary for transferred model
- Proposal 6 (Sec 4): Discuss how many model structures should be standardized for model transfer case z4
- Proposal 7 (Sec 4): Discuss the approaches to determine which model structures should be standardized. One practical approach is simulation evaluation with the calibration over companies
- Proposal 8 (Sec 4): Study the pros and cons of using the existing model format (e.g., ONNX) or introducing new 3GPP format
- Proposal 9 (Sec 4): Model ID can be defined by model structure + model parameter ID for model transfer/delivery case z4