R1-2410775
discussion
Summary #1 for other aspects of AI/ML model and data
From OPPO
Summary
This 3GPP RAN1 document (R1-2410775) from OPPO serves as a moderator summary for AI/ML model and data aspects in Rel-19, containing approximately 20+ proposals across model identification, training data collection, and model transfer/delivery. The document consolidates company contributions and proposes agreements for further study of controversial AI/ML topics including two-sided model identification options, dataset transfer mechanisms, and standardized model structures.
Position
OPPO, as the moderator, is advocating FOR a systematic study of model identification mechanisms for two-sided models, particularly supporting MI-Option2 with dataset transfer and standardized model structures for Case z4. They push FOR 3GPP specification of model structures rather than offline vendor collaboration, and advocate FOR comprehensive dataset information exchange. OPPO is positioning AGAINST overly complex model identification schemes that would burden implementation while supporting unified LCM frameworks that provide both functionality-based and model-ID-based operations.
Key proposals
- Proposal 2.1 (Sec 1): Study three alternatives for identifying UE-part of two-sided models: Alt. MI-2A using ID-X, Alt. MI-2B with network-assigned model ID, and Alt. MI-2C with UE-reported model ID
- Proposal 2.1A (Sec 1): Use ID-X for pairing UE-part and NW-part of two-sided models with further study on additional pairing information needed
- Proposal 2.2 (Sec 1): Transmit comprehensive dataset information including input data, labels, format/type details, dataset size, and validation/testing information for MI-Option2
- Proposal 2.3 (Sec 1): Study four options for ID-X definition: dataset ID, model ID, pairing ID, or associated ID for two-sided model identification
- Proposal 2.5 (Sec 1): Conclude that whether datasets include single or multi-cell data is network implementation and transparent to UE, with datasets potentially containing multi-cell and multi-condition data
- Proposal 2.6 (Sec 1): Study MI-Option4 with three cases - standardized reference model as UE part only, NW part only, or both parts of two-sided models
- Proposal 4.1 (Sec 3): Observe trade-offs between 3GPP-specified model structures (reduced cross-vendor burden but limited flexibility) versus offline NW-UE alignment (reduced standardization but increased collaboration burden)
- Proposal 4.2 (Sec 3): Focus RAN1 study on standardized known model structures for Case z4, noting offline alignment is beyond RAN1 expertise
- Proposal 4.4 (Sec 3): Study three open format options for model transfer: reusing existing formats like ONNX, defining new 3GPP format including ASN.1, or reusing SA2 interoperability tokens
- Proposal 4.5 (Sec 3): Consider multiple options for determining AI model readiness including UE signaling, specified minimum application time, or UE-reported minimum time with network configuration
- Proposal 4.7 (Sec 3): Study relationship between model ID, first indication, and second indication with three options: combined information, second indication as model ID, or network-assigned separate model ID
- Proposal 4.8 (Sec 3): Consider Transformer, CNN, and LSTM as starting point backbones for standardized model structures in Case z4