Age of Generative Information (AoGI): an Inference-Aware Metric for Freshness Evaluation in Real-Time Computer Vision
Xiao, Yuquan Du, Qinghe Cheng, Wenchi Karagiannidis, George K Nallanathan, Arumugam Guizani, Mohsen Wang, Jiangzhou
Xiao, Yuquan
Du, Qinghe
Cheng, Wenchi
Karagiannidis, George K
Nallanathan, Arumugam
Guizani, Mohsen
Wang, Jiangzhou
Supervisor
Department
Machine Learning
Embargo End Date
Type
Journal article
Date
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Generative Artificial Intelligence (GAI) is revolutionizing fashion in various industries at a rapid pace, especially in the field of real-time computer vision (CV). However, due to the high computational demands of GAI models, performing timely inference on local clients remains a significant challenge. This paper investigates the benefits of using a helper to assist the local client in accelerating inference, and poses a central question: towards diverse computational capabilities of devices and dynamic wireless fading environments, how should we design sampling, transmission, and model partitioning schemes to maintain the freshness of the generative information? To address this issue, we propose a novel metric called age of generative information (AoGI) to quantify the freshness of generative content. Specifically, AoGI measures the time elapsed since the prompt associated with the most recently completed generative information was sampled. Unlike the classical age of information (AoI), AoGI takes into account the effects of not only sampling and transmission, but also inference on information freshness. We then study the average peak AoGI minimization problem in dynamic wireless fading environments, which is fractional but nonconvex-concave. Using the Dinkelbach's transform as well as the majorization-minimization (MM) method, we pro pose an average-peak-AoGI-oriented (Avg-P-AoGI-O) sampling, transmission, and model partitioning scheme. In addition, for some extremely time-critical CV tasks, the maximum-peak AoGI minimization problem is studied. We solve it by using the MM method and propose a maximum-peak-AoGI-oriented (Max-P AoGI-O) scheme. Our results show that when the computational capabilities of the local client and the helper differ by less than a factor of two, model partitioning yields better AoGI performance. However, when the gap exceeds this threshold, offloading the entire model to the more powerful device is more effective. Numerical evaluations show that our proposed schemes can reduce the peak AoGI by about 25% compared to proportional model partitioning and inference delay-oriented baselines, thereby ensuring a higher freshness level of generative information.
Citation
Y. Xiao, Q. Du, W. Cheng, G.K. Karagiannidis, A. Nallanathan, M. Guizani , et al., "Age of Generative Information (AoGI): an Inference-Aware Metric for Freshness Evaluation in Real-Time Computer Vision," IEEE Transactions on Mobile Computing, vol. PP, no. 99, pp. 1-13, 2026, https://doi.org/10.1109/tmc.2026.3689106.
Source
IEEE Transactions on Mobile Computing
Conference
Keywords
40 Engineering, 4006 Communications Engineering, 46 Information and Computing Sciences
Subjects
Source
Publisher
IEEE