Abstract
Distributed machine learning at the network edge has emerged as a promising new paradigm. Various machine learning (ML) technologies will distill Artificial Intelligence (AI) from enormous mobile data to automate future wireless networking and a wide range of Internet-of-Things (IoT) applications. In distributed edge learning, multiple edge devices train a common learning model collaboratively without sending their raw data to a central server, which not only helps to preserve data privacy but also reduces network traffic. However, distributed edge training and edge inference typically still require extensive communications among devices and servers connected by wireless links. As a result, the salient features of wireless networks, including interference and channels' heterogeneity, time-variability, and unreliability, have significant impacts on the learning performance.