Seminars/Colloquia/Invited Talks

Seminars

Francois Fleuret

Conditional Independence for Prediction

PLACE:	Clark 314
EVENT:	CIS Seminar Series
DATE:	November 17, 2006
TIME:	1:00 - 2:00 PM

Abstract

In this talk I will present the simple and powerful idea of using conditional independence for prediction.
Modeling high-dimensional signals explicitly or through learning is often impossible due to a lack of either training data or computational power. However, a simple assumption of conditional independence of the signal components given the true state of interest or other hidden state variable often leads to a tractable form that can be handled with adequate algorithmic tricks.
I will first introduce this idea on a toy example, and then illustrate it with three different problems: face detection with a coarse-to-fine representation, multi-camera tracking with a probabilistic occupancy map, and learning from one sample with transfer learning.
In all these problems, I will also show the importance of algorithmic efficiency and how a tight integration between statistics and algorithmic design is the key to efficient and usable schemes.

Brief Biography:

Francois received a PhD in Probabilities from University of Paris VI in 2000 and then spent one year at the Department of Computer science at the University of Chicago. He then spent two years as a researcher at INRIA in France. After that he took a sabbatical leave to work in the CVLab at EPFL where he has been since 2004.