Edge Computing Solutions for Distributed Machine Learning
Fabrizio Marozzo, Alessio Orsino, Domenico Talia, Paolo Trunfio
Abstract
The rapid spread of the Internet of Things (IoT), with billions of connected devices, has generated huge amounts of data and asks for decentralized solutions for machine learning. However, performing complex learning tasks at the edge of the network is posing great challenges in terms of efficient management of data storing, transfer, and analysis. For these reasons, a lot of research and development effort is devoted to adapt different machine learning algorithms so that cooperative training and inference on local data occur directly at the edge of the network. This scenario represents a major challenge today due to the limited capacities of edge devices, the different technologies with which these devices work and communicate, and the lack of common software stacks to easily manage them. In this paper, we analyze distributed machine learning algorithms and how they should be adapted to run at the network edge and, if needed, cooperate with the cloud to ensure low latency, energy savings, privacy preserving and scalability. In particular, we briefly discuss how the main machine learning algorithms have been adapted to work in traditional distributed platforms (such as clusters, clouds, and HPC systems) and the main research work that has led these algorithms to run on resource-constrained edge devices. Then, a layered approach is introduced and discussed for adapting machine learning algorithms on edge-cloud architectures. Finally, we conclude the paper by describing some application scenarios that can benefit from this approach.