Developing a machine learning library for microcontrollers

Abstract

With the appearance of tools to support the emerging paradigm of edge computing, we expect that low cost microcontrollers will become appealing execution platforms also for machine learning. To explore this field, we implemented Machine Learning eMbedded Library (ML)2 and tested it in a simple case (classifying human movement as normal or not) and with a benchmark dataset to have a first comparison in performance with other implemented algorithms. Results—in terms of accuracy and of execution time, both for training and classification—are promising, and encourage the next steps of our work, in the direction of extending the set of implemented algorithms and going more in depth with the testing. In any case, we believe that these preliminary results should spur the Internet of Things research community in devising distributed computing algorithms able to support ML computation as close as possible to the source.