APPLIED MATH SEMINAR
Speaker: Boaz Nadler, Weizmann Institute of Science

Title: Uncertainty Estimates, Classification and Active Learning

When/where: Tuesday, December 7th, 4:15 PM, LOM 200

Abstract: In contrast to the theoretical framework of classification, where the goal
is to construct a classifier with small risk, in many real world applications
simple predictions of class labels or even point estimates for posterior class
probabilities are typically insufficient. These quantities may
be highly inaccurate for test samples not well represented by the training set, a
common situation when the training set is small or when training and test data
have different distributions. Instead, some measure of confidence in these point
estimates is required as well.
In this talk we propose a Bayesian framework to derive such confidence measures for several popular classifiers, including epsilon-NN, k-NN and random forest. We illustrate our approach, which unifies classification and outlier detection, on several datasets. Finally, we show its utility for active learning (AL). We derive a novel uncertainty based AL scheme and show its superior
performance in comparison to several competitors.
Joint work with Jens Roeder, Michael Hanselmann and Fred Hamprecht (Heidelberg University).