Towards Robust and Secure Machine Intelligence: A Data Fusion Perspective
We dream of living in a smart world that seamlessly blends people, computers, machines, infrastructure, and information. As technical capability for enabling superior levels of machine intelligence is getting closer than ever before, world leaders in computing, machine learning and artificial intelligence are now raising serious concerns about the danger of nextgen intelligent machines behaving in unpredictable, erratic and potentially hazardous manners. These concerns appear to be centered around: (a) the unprecedented power that can be acquired by intelligent machines by merely learning from data, and (b) our inability understand what’s actually being learned (especially in light system response to unforeseen data) due to the complexity of deep hierarchical structures that enable this unparalleled learning capacity. On the other hand, the traditional approaches that allows for complete characterization of system behavior, for instance via state-space models significantly lack in representational complexity required for machine intelligence. As engineers, we must be able to build systems that behave in a robust, secure and predictable manner, especially for those machines that are responsible for our healthcare, energy, defense, basic infrastructure, and other aspects governing the safety and welfare of human beings. First, how can we build machines so that their behavior to unforeseen data is predictable? Second, how can we make use of the unparalleled learning capacity of advanced data-driven methods for enabling intelligent behaviors while making sure that they function in a predictable manner on the face of unforeseen data? In this talk, some first steps towards achieving robust and secure machine intelligence is presented from a data fusion perspective. In particular, via examples derived from on-going work on SAFENETS, a Situational Awareness Framework for Emerging Network Enabled Transportation Systems, synergistic use of advanced knowledge extraction, data modeling, integrated sensing, signal processing, and machine learning towards achieving robust system behavior via a systematic approach is discussed. The presented work is relevant to many domains, including but not limited to advanced manufacturing, healthcare, smart infrastructure, and defense.
Thanuka Wickramarathne is an Assistant Professor with the Department of Electrical and Computer Engineering at the University of Massachusetts Lowell (UMass Lowell), having joined the faculty in Fall'16. Prior to joining UMass Lowell, he was a Research Assistant Professor at the University of Notre Dame in the Depts. of Electrical Engineering, and Computer Science and Engineering. He received his B.Sc. in Electronics and Telecommunication Engineering from the University of Moratuwa, Sri Lanka and both his M.S. and Ph.D. degrees from the University of Miami, specializing in signal processing with a focus on data fusion, machine learning and knowledge extraction in imperfect data domains. His current research interests are in the broader areas of multi-sensor data fusion, with a particular emphasis on machine intelligence, big-data signal processing and situational awareness with ubiquitous sensing.
Host: Professor Alex Wyglinski