R1-2410022
discussion
Discussion on other aspects of AI/ML model and data
From Lenovo
Summary
Lenovo's document discusses model identification procedures for two-sided AI/ML models in CSI compression, focusing on inter-vendor training collaboration options. The document contains 9 key proposals addressing model structure sharing, dataset identification, and standardized model formats for collaborative AI/ML implementation across different vendors.
Position
Lenovo advocates FOR standardized model identification procedures that enable inter-vendor collaboration while maintaining vendor-specific optimization capabilities. They support using dataset IDs (ID-X) for model-dataset association and favor exploring multiple open format options including existing AI community formats (ONNX) and 3GPP-defined formats. Lenovo pushes FOR deferring Option 5 model identification decisions until model format standardization consensus is reached.
Key proposals
- Proposal (Sec Model identification Option 3): Consider indicating the supported/candidate model structures to the other side during the model identification procedure
- Proposal (Sec Model identification Option 4): To use ID-X for the dataset to assist exchanging the information on the models developed, and the relation between ID-X and data collection related configuration(s) can be further studied
- Proposal (Sec Model identification Option 5): Study necessity of model identification for the inter-vendor training collaboration Option 5 after there is consensus on the model format standardization
- Proposal (Sec Conclusion): Study the relevant information/configurations on data collection, dataset and model transfer, to be shared during model identification procedure for the two-sided model of the AI/ML-based CSI compression use case
- Proposal 4.3 Option 1 (Sec Proposal 4.3): Reuse the existing open format(s) that has existed in the AI community (e.g., ONNX)
- Proposal 4.3 Option 2 (Sec Proposal 4.3): Define a new open format within 3GPP
- Proposal 4.3 Option 3 (Sec Proposal 4.3): Using ASN.1 to represent the AI model
- Proposal 4.3 Option 4 (Sec Proposal 4.3): Reuse the mechanism defined in SA2 (interoperability token) for aligning model description format