Network AI Management & Orchestration - A Federated Multi-task Learning Case.

Rongpeng Li,Wenliang Liang,Chenghui Peng,Xueli An,Zhifeng Zhao,Honggang Zhang
DOI: https://doi.org/10.1109/GCWkshps52748.2021.9681969
2021-01-01
Abstract:6G treats artificial intelligence (AI) as the corner-stone and fundamental paradigm shift for providing inclusive intelligent services, which requires to natively support the training and reasoning of AI and provide a comprehensive network AI management & orchestration (NAMO) solution. However, NAMO faces many practical challenges like multi-tenant multi-task coordination, heterogeneous resource scheduling, and security & privacy concerns. In this paper, we take the federated multi-task learning as a starting case to demonstrate a promising NAMO solution. In particular, we propose a resource-aware method which leverages a primal-dual relationship to allow no direct up-loading of local data to the edge server and maintain synchronous updates with straggler tolerance. Also, the proposed method could dynamically tune the learning accuracy at devices and the number of federated iterations to obtain a satisfactory training accuracy. Extensive simulation results have demonstrated the effectiveness of the proposed method.
What problem does this paper attempt to address?