Discriminative Unsupervised Learning of Structured Predictors
- Linli Xu
- Dana Wilkinson, School of Computer Science, University of Waterloo
- Finnegan Southey
- Dale Schuurmans, AICML
We present a new unsupervised algorithm for training structured predictors that is dis- criminative, convex, and avoids the use of EM. The idea is to formulate an unsuper- vised version of structured learning meth- ods, such as maximum margin Markov net- works, that can be trained via semidenite programming. The result is a discrimina- tive training criterion for structured predic- tors (like hidden Markov models) that re- mains unsupervised and does not create lo- cal minima. To reduce training cost, we reformulate the training procedure to mit- igate the dependence on semidenite pro- gramming, and nally propose a heuristic procedure that avoids semidenite program- ming entirely. Experimental results show that the convex discriminative procedure can produce better conditional models than con- ventional Baum-Welch (EM) training.
Citation
L. Xu, D. Wilkinson, F. Southey, D. Schuurmans. "Discriminative Unsupervised Learning of Structured Predictors". International Conference on Machine Learning (ICML), Pittsburgh, January 2006.Keywords: | discriminative, structured, predictors, machine learning |
Category: | In Conference |
BibTeX
@incollection{Xu+al:ICML06, author = {Linli Xu and Dana Wilkinson and Finnegan Southey and Dale Schuurmans}, title = {Discriminative Unsupervised Learning of Structured Predictors}, booktitle = {International Conference on Machine Learning (ICML)}, year = 2006, }Last Updated: April 24, 2007
Submitted by AICML Admin Assistant