• Calendar

December 14, Tuesday
12:00 – 13:30

I will introduce confidence-weighted linear classifiers, a class of algorithms that maintain confidence information about classifier parameters. Instead of a single weight vector, learned hypotheses are given by Gaussian distributions over weight vectors, with a covariance matrix that represents uncertainty about weights and correlations between different weights. Learning in this framework updates parameters by estimating weights and increasing model confidence. I will describe few online algorithms that maintain a Gaussian distribution over weight vectors, updating the mean and variance of the model with each instance. A mistake bound analysis shows that indeed the algorithm performs better under some conditions and also relates between our model and the margin and loss analysis of previous models. Empirical evaluation on a range of NLP tasks show that our algorithm improves over other state of the art online and batch methods, learns faster in the online setting, ends itself to better classifier combination after parallel training and is suites better for active learning. Based on joint work with Mark Dredze,Alex Kulesza and Fernando Pereira.