Speaker : Professor Laurent Younes(Johns Hopkins University)
Title : Learning Multivariate Distributions by Competitive Assembly of Marginals
Time : 2013-01-11 (Fri) 14:30 -
Place : Seminar Room 722, Institute of Mathematics (NTU Campus)
Abstract: We present a two-step framework for learning high-dimensional multivariate probability distributions. The first step learns and selects statistical building blocks, or “primitives”, which are low-dimensional marginal distributions on small subsets of variables. The second step is a combinatorial optimization that computes an optimal assembly of building blocks subject to consistency constraints. Only a small fraction of the primitives participates in any valid construction. The whole approach is based on a paradigm in which complex models are built by progressively merging previously validated lower-dimensional marginal distributions. Performance is evaluated using both synthetic data and real datasets from natural language processing and computational biology. Joint work with F. Sanchez-Vega, J. Eisner and D. Geman