On Deploying Machine Learners into Embedded Systems


Tommaso Zoppi, Andrea Ceccarelli, Andrea Bondavalli
University of Florence

Presentation title

On Deploying Machine Learners into Embedded Systems

Authors

Tommaso Zoppi
University of Florence
Andrea Ceccarelli
University of Florence
Andrea Bondavalli
University of Florence

Presentation type

Technical presentation

Abstract

Machine Learning (ML) algorithms became the enabling condition of many classification, optimization and regression tasks that aim at fulfilling a wide variety of tasks in ICT systems. Particularly, they may be integrated into embedded systems for the purpose of load balancing, image processing, error detection, intrusion detection or failure prediction, to name a few. Those ML algorithms require data and time for learning their model, which also has to be stored when deployed into a real system. However, devices and sensors to be equipped with ML algorithms usually have tight constraints both in space and time which could be dedicated to those detection, prediction or optimization tasks, and also aim at performing very accurate tasks. Therefore, this presentation shows how different supervised and unsupervised ML algorithms have different accuracy, time and space complexity for performing different tasks. The discussion is carried out by means of plots, graphs and practical examples, showing the methodology to be used to derive the optimal ML algorithm to be installed in a device or a sensor to perform classification, optimization or regression tasks.


Additional material

  • Presentation slides: [pdf]

For more details on this presentation please click the button below: