[PAST EVENT] Colloquium: Queries to Bandits - Learning by Interacting

February 10, 2012
8am - 8:50am
McGlothlin-Street Hall
251 Jamestown Rd
Williamsburg, VA 23185Map this location
Machine learning has revolutionized virtually every domain where data is abundant. This is no surprise -- the key insight of machine learning is that generic algorithms that learn and evolve as they see data often outperform specialized methods painstakingly devised by experts. Training a machine learning algorithm, however, is often expensive because determining labels for training data can be costly. Fortunately, in many settings, learning algorithms are able to decide which data to have labeled or to otherwise interact with their environment. This talk will focus on understanding and exploiting the power of this interaction.

We will consider work in two domains. First, we will discuss learning interaction networks: finite populations of elements whose state may change as a result of interacting with other elements according to specific rules. These include a variety of structures, from social organizations to gene regulatory networks. We will see how to applying the query learning framework to these structures gives rise to practical models, deep theoretical questions and interesting algorithms.

Then, we will turn to the classical multi-armed bandit problem, which essentially captures many situations from medical testing to computational advertising. In the bandit framework, the learner's feedback is again guided by its own decisions, and it must carefully balance exploiting learned strategies and exploring new ones. We will examine some recent work in this area, which has resolved fundamental open questions and indicated the possibility of deploying optimal bandit algorithms in practice.

Lev Reyzin is a postdoctoral fellow at the Algorithms and Randomness Center at the Georgia Institute of Technology. His research interests span the theory and practice of machine learning. Before coming to Georgia Tech, he spent a year as a scientist at Yahoo! Research working on machine learning problems in computational advertising. After earning a B.S.E. from Princeton University, Lev completed his Ph.D. in 2009 at Yale University; his dissertation work focused on active learning of interaction networks. His work has received awards at ICML, COLT and AISTATS, as well as an NSF Graduate Fellowship, an NSF Computing Innovation Fellowship and a Simons Postdoctoral Fellowship.

Department of Computer Science