Abstract: Imagine living in an intuitively interactive space without the need to understand its grammar. One doesn't need to interact in a specific way or use voice commands, like Amazon echo, or always wear something, like a smart band to control the space and the devices therein. Sharing this intelligent space with others does not also degrade the individual user experience. Interestingly, this vision of seamless smart spaces is not novel and quite dated; however, we are yet to occupy this kind of space regularly. For this vision to become an everyday reality, we argue that there is a need for multi-user continuous activity tracking through passive sensing mechanisms. In this talk, I'll discuss how we have developed a pervasive sensing system for continuous multi-user room-scale tracking of human activities using a single COTS mmWave radar. We'll further highlight how such a sensing system can be utilized for continuous monitoring of driving activities while on the road, to prevent road accidents from driver distraction and thus to promote safe driving. We'll conclude the talk with some potential research challenges where mmWave sensing can help.

Abstract: Deep neural networks have rapidly grown in their ability to address complex learning problems. Consequently, they are being integrated into society with an ever-rising influence on all levels of human experience. This has resulted in a need to gain human-understandable insights in their internal mechanisms to ensure they operate ethically and reliably. The study and development of methods which can generate such insights constitutes the field of interpretable machine learning. The talk will focus on the use of dictionary learning methods to discover relevant patterns inside hidden layers of deep networks. Specifically, we will explore their application for post-hoc interpretation in two different styles and contexts: (1) Generating listenable interpretations for audio processing networks, (2) Understanding internal representations of large multimodal models.