The explosive growth in digital archiving poses many fundamental challenges in machine learning. This talk will describe a confluence of recent ideas from applied mathematics, statistics, and machine learning on constructing low-dimensional representations that capture the intrinsic statistical properties of the original data. I will present new representation discovery algorithms motivated by two application domains: cross-lingual information retrieval and robotics. I will describe new alignment algorithms that find correspondences across languages by projecting document collections onto a common low-dimensional manifold. I also describe new reinforcement learning algorithms for solving robot control problems modeled as Markov decision processes that dynamically construct sparse representations for approximating value functions.
BIO: Sridhar Mahadevan is a professor in the Computer Science Department at the University of Massachusetts, Amherst, where he co-directs the Autonomous Learning Laboratory. Professor Mahadevan's research interests span several areas of artificial intelligence and computer science, including machine learning, decision making, and robotics. He has published over a hundred articles in these areas, as well as three books.