Tuning Vision Transformer with Device-to-Device Communication for Targeted Image Recognition
Hideya Ochiai, Atsuya Muramatsu, Yudai Ueda, Ryuhei Yamaguchi, Kazuhiro Katoh, Hiroshi Esaki · 2023
Edge machine learning is now applied to many industrial sectors. The captured data in their business is for their mission-oriented tasks and is originally collected at multiple edge devices locally. This paper explores the expansions of edge-based collaborative training - without relying on communication infrastructures such as 5G and the Internet as an alternative to legacy machine learning schemes. This paper proposes the use of a Vision Transformer for image recognition tasks in device-to-device collaborative scenarios inspired by wireless ad hoc federated learning (WAFL). We have developed UTokyo Building Recognition Dataset and evaluated the performance of our proposed model. The results indicate that our proposed model outperforms WAFL-based other deep learning models such as WAFL-ResNet, VGG, and MobileNet with less device-to-device communication costs.