EFFICIENT AND SCALABLE MACHINE LEARNING FOR DISTRIBUTED EDGE INTELLIGENCE

Abstract

In the era of big data and IoT, devices at the edges are becoming increasingly intelligent, and processing the data closest to the sources is paramount. However, conventional machine learning works with the centralized framework of collecting data from various edge sources and storing it on the high-performance cloud to support computationally intensive iterative algorithms. As a result, such a framework is infeasible for training on embedded edge devices with limited resources and a tight power budget. This dissertation proposes to integrate ideas from fields of machine learning and systems for designing efficient and scalable algorithms for distributed training of machine learning models amenable for edge computing with limited hardware and computing resources. The resulting decentralized machine learning framework aims to keep the data private, reduce latency, save communication bandwidth, be energy-efficient, and handle streaming data.