On-device Sensing
TL;DR: Presentation on on-device sensing to fulfill intelligent tasks, current system challenges and solutions.
Abstract:
Machine learning and connected devices are causing a cambrian explosion of commercial and research applications that can sense the world. From cameras to motion sensors and microphones, intelligent apps require some kind of inference (and increasingly training) with deep neural networks, which however can easily overwhelm the resources of constrained platforms. In this talk we will discuss recent research, system challenges, and solutions. From meta-learning and distilation, to self-supervised and semi-supervised learning (to name a few), here we will review the state of the art in on-device sensing for activity recognition towards lightweight and privacy-aware ubiquitous learning.