Delay Analysis of Wireless Federated Learning Based on Saddle Point Approximation and Large Deviation Theory

• DOI: 10.48550/arxiv.2103.16994
• Publication Date: 2021-01-01
• Authors: Lintao Li; Longwei Yang; Xin Guo; Yuanming Shi; Haiming Wang; Wei Chen; Khaled B. Letaief
• Keywords: Saddle point; Saddle

Federated learning (FL) is a collaborative machine paradigm, which enables deep model training over large volume of decentralized data residing in mobile devices without accessing clients' private data. Driven by the ever increasing demand for applications or devices, vast majority FL tasks are implemented wireless fading channels. Due to time-varying nature channels, however, random delay occurs both uplink and downlink transmissions FL. How analyze overall time consumption task, more specifically, FL's distribution, becomes challenging but important open problem, especially delay-sensitive training. In this paper, we present unified framework calculate approximate distributions arbitrary Specifically, saddle point approximation, extreme value theory (EVT), deviation (LDT) jointly exploited find distribution along with its tail characterizes quality-of-service system. Simulation results will demonstrate that our approximation method achieves small error, vanishes increase accuracy.