Speaker : Dr. Lo Bin Chang (Brown University)
Title : Conditional modeling and conditional inference
Time : 2011-02-10 (Thu) 15:00 - 16:00
Place : Lecture Hall, Inst. of Mathematics
Abstract: The goal is to construct or test probability models for complex systems, such as images and image interpretations or stock and derivative prices and volatility. I will give a variety of examples in which a high dimensional joint distribution can be effectively modeled or analyzed by focusing attention either on a low-dimensional marginal distribution or a low-dimensional conditional distribution given a high-dimensional marginal. The idea is to first select an aspect of the distribution that is of particular interest, or is particularly in conflict with the data, and then to manage its modeling or analysis in a way that is “agnostic” (invariant) to the distribution of the remaining degrees of freedom. I will discuss applications to building non-Markovian hierarchical models for scene analysis, appearance models for objects and object parts, efficient computation and near-optimal ROC performance through sequential testing, and the statistical analysis of Black-Scholes models and their extensions.