R1-2409672
discussion
Other aspects of AI/ML model and data
From vivo
Summary
This document from vivo analyzes model identification and transfer mechanisms for NR AI/ML, specifically focusing on the feasibility of Case z4 (known model structure transfer). It presents 22 proposals and 7 observations covering associated ID scopes, reference model standardization, hyper-parameter specifications for CNN/Transformer backbones, and signaling procedures for parameter updates.
Position
vivo proposes supporting local associated IDs for multiple cells to balance training consistency with network privacy, rather than using global IDs that expose deployment choices. They conclude that Case z4 (known model structure transfer) is feasible from a device implementation perspective, citing lab tests showing low latency for parameter updates. vivo requires the specification of detailed hyper-parameters for standardized CNN and Transformer backbones, including kernel sizes, strides, and attention head dimensions. They propose using ASN.1 as the starting point for parameter signaling and define a two-part indication mechanism (structure + parameters) to manage model identification. Additionally, they argue that partial parameter transfer and UE reporting of parameter readiness are necessary to reduce overhead and ensure synchronization.
Key proposals
- Proposal 1 (Sec 2.1.1): Supports local associated ID for multiple cells to maintain training/inference consistency while exposing fewer network deployment choices than global IDs.
- Proposal 2 (Sec 2.1.1): Proposes supporting Associated ID combined with cell ID(s) to indicate applicable cells in multi-cell scenarios.
- Proposal 4 (Sec 2.1.3): Proposes supporting Associated ID with timestamp information to define the applicable period of the ID.
- Proposal 5 (Sec 2.2): Proposes further studying Dataset ID, Associated ID, and Model ID as options for 'ID-X' in dataset transfer (MI-Option 2).
- Proposal 6 (Sec 2.3): States that model identification is needed when multiple models are transferred from NW to UE to enable low-latency model switching.
- Proposal 8 (Sec 2.4): Proposes two procedures for MI-Option 4: UE reporting global model ID directly, or NW assigning a local model ID after UE reports the global ID.
- Proposal 11 (Sec 3.1): Concludes that model transfer in open format of a known model structure (Case z4) is feasible from a device implementation perspective.
- Proposal 13 (Sec 3.2.1): Proposes a three-step procedure to align reference model structures: determine backbone/hyper-parameters, align evaluation assumptions, and align hyper-parameters based on results.
- Proposal 14 (Sec 3.2.2): Specifies that for CNN backbones, standardized structures must define activation functions and convolution hyper-parameters (kernel size, stride, padding).
- Proposal 15 (Sec 3.2.3): Specifies that for Transformer backbones, standardized structures must define block count, embedding length, self-attention dimensions, and feed-forward block parameters.
- Proposal 16 (Sec 3.3): Proposes using ASN.1 signaling as the starting point for model parameter exchange over the air interface.
- Proposal 17 (Sec 3.4): Concludes that UE must report whether it needs new parameters for a known model structure to reduce signaling overhead.
- Proposal 19 (Sec 3.4): Concludes that NW could indicate the transmission of partial model parameters to optimize overhead.
- Proposal 20 (Sec 3.4): Proposes referring to a model by the combination of a first indication (model structure) and a second indication (transferred parameters).