Surrogate Data for Deep Learning Architectures in Rehabilitative Edge Systems

Abstract

The age of Big Data came about with the profitable mining of large amounts of data stored on the Cloud data servers accessible through the Internet. Much of these data were provided by users of extensive social networks. The availability of these data for training, coupled with advances in processing power have led to the surge in deep learning applications.While the Cloud provides wide scale data storage and analytic facilities, there is a move to perform data analyses closer to where data is gathered. Low resource but powerful processors have led the move toward these systems at the “edge“ rather than remotely.However, certain environments which can benefit from edge systems do not produce large volumes of data, such as in clinical applications. We augment such data using surrogate time series and compare various neural network architectures which would allow data analysis in rehabilitative edge systems.We show that a neural network using temporal data provides excellent classification results while using a fraction of the resources in an earlier work. Our novel framework shows how deep learning tools can be trained in a data scarce environment and then deployed in resource constrained edge systems.