Age of Model: A Metric for Keeping AI Models Up-to-Date in Edge Intelligence-Enabled IoV
Xiaoyu Chi, Xiaoqi Qin, Xiaodong Xu, Shujun Han, Liang Jin, Ping Zhang · IEEE Transactions on Vehicular Technology · 2024
In the 6th generation (6G) mobile networks, intelligence is expected to be present in every part of the Internet of Vehicle (IoV) system to support smart transportation. In the context of constantly changing task requirements caused by dynamic environments of IoV, lifelong model update is essential to maintain high-quality inference performance. Given the mismatch between the resource demanding update process and limited system resource, it is critical to strike a balance between update quality and the update frequency. In this paper, we propose a performance metric called Age of Model (AoM) to characterize the inference performance degradation of AI models under dynamic task requirements. We derive the closedform expressions of AoM at both edge tier and cloud tier, where various queues are exploited to abstract the model update process with single server or multiple servers. The obtained analytical framework describes the dependency among update request frequency, computing rate and training data diversity, which can serve as performance metric for system optimization. The accuracy and feasibility of our analysis are verified with extensive simulation results.