Decentralized Descent Optimization With Stochastic Gradient Signs for Device-to-Device Networks

Abstract

We propose an algorithm for decentralized optimization in wireless device-to-device (D2D) networks of pervasive devices such as sensors or 5G handsets, in which the signs of stochastic gradient are used for descent steps. Our algorithm has the convergence rate of ${O}$ (1/( nT )) in which ${n}$ is the number of devices and ${T}$ is the number of learning iterations, saving the communication efficiency by at least 64 times when compared with previous results, and being relatively robust to unexpected errors of adversarial scaling in communication. Theoretical claims are verified by numerical results on a standard benchmark dataset.