EDT-Edge: An Elastic Distributed Training Framework for 6G Edge Networks
Feng, Wenjiao Xiao, Rongxing Li, Zonghang Sun, Gang Yu, Hongfang Niyato, Dusit
Feng, Wenjiao
Xiao, Rongxing
Li, Zonghang
Sun, Gang
Yu, Hongfang
Niyato, Dusit
Supervisor
Department
Machine Learning
Embargo End Date
Type
Conference proceeding
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Edge Intelligence (EI) has become a promising paradigm for integrating Artificial Intelligence (AI) technologies with the edge network for low-latency service provisions in future 6G scenarios. However, it is challenging to implement AI model training(AI-MT) at the 6G network edge due to limited communication and computing resources, as well as dynamic network topology. To address these challenges, we propose EDT-Edge, an elastic distributed training framework for efficient AI-MT at the 6G network edge with flexibility, reconfiguration, and scalability advantages. Specifically, our EDT-Edge enables adaptive topology adjustments and concurrent single-hop fine-grained replication of training states, which ensures seamless training cooperation among multiple edge servers. The experimental results show that our EDT-Edge reduces computing resource consumption by 50% and 86% compared to the state-of-the-art Stop-Free and Stop-Resume manners, respectively. Meanwhile, our EDT-Edge decreases training states' replication delay by 53% and 88% compared to the SPT and Single-Node schemes, respectively.
Citation
W. Feng, R. Xiao, Z. Li, G. Sun, H. Yu and D. Niyato, "EDT-Edge: An Elastic Distributed Training Framework for 6G Edge Networks," IEEE INFOCOM 2025 - IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS), London, United Kingdom, 2025, pp. 1-6, doi: 10.1109/INFOCOMWKSHPS65812.2025.11152954.
Source
Proceedings of the IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS)
Conference
2025 IEEE Conference on Computer Communications Workshops, INFOCOM WKSHPS 2025
Keywords
6G, Edge Intelligence, elastic distributed training framework, seamless training
Subjects
Source
2025 IEEE Conference on Computer Communications Workshops, INFOCOM WKSHPS 2025
Publisher
IEEE