Making Mobiles Do More with Less through Approximate Computing

Mobile computing proliferation is critically threatened by the breakdown of Dennard scaling, a law describing the area-proportional growth of integrated circuit power use. As a consequence, further transistor packing in a mobile device will require additional energy at times where the increased battery life is already the most desired property that mobile users would like to see. In our work we propose Approximate Mobile Computing (AMC) and take a radical stance that computation does not need to be 100% precise. We first examine situations, such as mobile video playback and mobile deep learning for human activity recognition, where the properties of the input and the limitations of human perception open space for AMC. We then develop methods that bring AMC to consumer devices. Unique to our work is the development of execution adaptation algorithms that dynamically tune the level of approximation according to the context of usage. Finally, we will discuss future steps towards making AMC widespread and easy to deploy.

Biography:
Veljko Pejović received his PhD in computer science from the University of California Santa Barbara, USA. From 2015 he is an assistant professor at the Faculty of Computer and Information Science (UL FRI), University of Ljubljana, Slovenia. Prior to this, he was a Research Fellow at the Computer Science Department, University of Birmingham, UK. At UL FRI he is leading research on mobile computing, focusing on resource-efficient mobile systems, human-computer interaction, and cybersecurity in ubiquitous systems. His awards include the best paper nomination at ACM UbiComp and the first prize at Orange D4D challenge for his work on epidemics modeling. More about his research can be found at http://lrss.fri.uni-lj.si/Veljko/