Events Calendar

Modeling and Tracking Activities with Event-Coupled HMMs

Speaker: Dr. Aram Galstyan, ISI, USC
Host: Prof. Jonathan GratchAbstract:
Plan, activity, and intent recognition (PAIR) is concerned with inferring hidden states of agents based on an observable sequence of their actions. Although PAIR has been an active area of research for more than a decade, most studies so far has been limited to systems with a single agent, or a handful of them. I will present our work on activity recognition on a larger scale, where thousands of agents interact with each other by engaging in abstract "attribute trades". Those interactions induce an evolving network, where the nodes and the edges represent the agents and their transactions, respectively. The collective dynamics of this network can be naturally modeled through a novel type of interacting hidden Markov models (HMM), which we call Event-Coupled HMMs. I will discuss our approach to scalable inference-making with EC-HMM, which involves pruning the network through semi-supervised learning, and utilizing an approximate and scalable representation of the hidden process on the reduced network. I will conclude by discussing the notion of "trackability", which can be intuitively defined as one's ability to accurately infer stochastic processes, and present some recent results in the context of binary HMMs.Biography:
Dr. Aram Galstyan received his Ph.D. in theoretical condensed matter physics from University of Utah, in 2000. He then joined USC Information Sciences Institute where he currently works as a computer scientist at the Intelligent Systems Division. Dr. Galstyan's current research focuses on learning and discovering patterns in large-scale sequential data, statistical network analysis, and semi-supervised learning with graphs. His other research interests include mathematical modeling of complex adaptive systems, emergent coordination in robotic swarms, and learning in multi-agent systems.

Location: Von Kleinsmid Center For International & Public Affairs (VKC) - 210

Audiences: Everyone Is Invited

Contact: CS Colloquia