@article{M940FE8A7,
title = "Multi-Exit Faster R-CNN(MEF): An Adaptive Multi-Exit Neural Network for Time-Varying Computing Resources",
journal = "The Journal of Korean Institute of Communications and Information Sciences",
year = "2025",
issn = "1226-4717",
doi = "10.7840/kics.2025.50.2.234",
author = "Hyeonsu Kim, Dongwoo Goo, Dayeon Lee, Minseok Choi",
keywords = "Object detection, Multi-exit neural network, Lyapunov optimization",
abstract = "This paper develops a multi-exit Faster R-CNN model and an exit selection algorithm that can adaptively adjust accuracy and latency in a fluctuating resource environment. Despite the rapid advancements in computer vision and natural language processing, where inference latency is increasingly critical, existing studies lack models and algorithms that ensure an optimal trade-off between latency and accuracy in dynamically changing environments. This study proposes a multi-exit neural network technique and a Faster R-CNN model utilizing Lyapunov optimization to address these issues. The proposed approach guarantees long-term optimal performance while maintaining system stability under uncertain resource changes and allows for the selection of the optimal exit for inference based on resource conditions. Additionally, simulations have verified that the proposed technology can adjust the trade-off between accuracy and latency according to available resources."
}