Exploiting the edge power: an edge deep learning framework

Abstract

Crowdsourced big data from Internet users have long been of interest to modern machine learning technologies, and recent advances in deep learning have shown great potentials in exploring the hidden information therein. Deep learning relies on strong computation power to process the massive amount of data, which is typical offered by modern data centers, so for data storage. A cloud built on top of the data center, which seamlessly integrates storage and computation, seems to be an ideal platform for learning. It however faces significant challenges from data collection and service distribution over the network, given the end users are globally and remotely distributed. In this article, we present edge learning for networked intelligent applications, which complements the cloud-centric design to effectively reduce network traffic and inference latency. We discuss the key design issues of edge learning, including strategies to push data pre-processing and preliminary learning to the network edge, as well as to confine computation to local regions with high accuracy. A prototype demonstrates its feasibility with off-the-shelf hardware and confirms its superiority with realworld experiments.