Learning Fast-Mixing Models for Structured Prediction

Jacob Steinhardt, Percy Liang
Proceedings of the 32nd International Conference on Machine Learning, PMLR 37:1063-1072, 2015.

Abstract

Markov Chain Monte Carlo (MCMC) algorithms are often used for approximate inference inside learning, but their slow mixing can be difficult to diagnose and the resulting approximate gradients can seriously degrade learning. To alleviate these issues, we define a new model family using strong Doeblin Markov chains, whose mixing times can be precisely controlled by a parameter. We also develop an algorithm to learn such models, which involves maximizing the data likelihood under the induced stationary distribution of these chains. We show empirical improvements on two challenging inference tasks.

Cite this Paper


BibTeX
@InProceedings{pmlr-v37-steinhardtb15, title = {Learning Fast-Mixing Models for Structured Prediction}, author = {Steinhardt, Jacob and Liang, Percy}, booktitle = {Proceedings of the 32nd International Conference on Machine Learning}, pages = {1063--1072}, year = {2015}, editor = {Bach, Francis and Blei, David}, volume = {37}, series = {Proceedings of Machine Learning Research}, address = {Lille, France}, month = {07--09 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v37/steinhardtb15.pdf}, url = {https://proceedings.mlr.press/v37/steinhardtb15.html}, abstract = {Markov Chain Monte Carlo (MCMC) algorithms are often used for approximate inference inside learning, but their slow mixing can be difficult to diagnose and the resulting approximate gradients can seriously degrade learning. To alleviate these issues, we define a new model family using strong Doeblin Markov chains, whose mixing times can be precisely controlled by a parameter. We also develop an algorithm to learn such models, which involves maximizing the data likelihood under the induced stationary distribution of these chains. We show empirical improvements on two challenging inference tasks.} }
Endnote
%0 Conference Paper %T Learning Fast-Mixing Models for Structured Prediction %A Jacob Steinhardt %A Percy Liang %B Proceedings of the 32nd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2015 %E Francis Bach %E David Blei %F pmlr-v37-steinhardtb15 %I PMLR %P 1063--1072 %U https://proceedings.mlr.press/v37/steinhardtb15.html %V 37 %X Markov Chain Monte Carlo (MCMC) algorithms are often used for approximate inference inside learning, but their slow mixing can be difficult to diagnose and the resulting approximate gradients can seriously degrade learning. To alleviate these issues, we define a new model family using strong Doeblin Markov chains, whose mixing times can be precisely controlled by a parameter. We also develop an algorithm to learn such models, which involves maximizing the data likelihood under the induced stationary distribution of these chains. We show empirical improvements on two challenging inference tasks.
RIS
TY - CPAPER TI - Learning Fast-Mixing Models for Structured Prediction AU - Jacob Steinhardt AU - Percy Liang BT - Proceedings of the 32nd International Conference on Machine Learning DA - 2015/06/01 ED - Francis Bach ED - David Blei ID - pmlr-v37-steinhardtb15 PB - PMLR DP - Proceedings of Machine Learning Research VL - 37 SP - 1063 EP - 1072 L1 - http://proceedings.mlr.press/v37/steinhardtb15.pdf UR - https://proceedings.mlr.press/v37/steinhardtb15.html AB - Markov Chain Monte Carlo (MCMC) algorithms are often used for approximate inference inside learning, but their slow mixing can be difficult to diagnose and the resulting approximate gradients can seriously degrade learning. To alleviate these issues, we define a new model family using strong Doeblin Markov chains, whose mixing times can be precisely controlled by a parameter. We also develop an algorithm to learn such models, which involves maximizing the data likelihood under the induced stationary distribution of these chains. We show empirical improvements on two challenging inference tasks. ER -
APA
Steinhardt, J. & Liang, P.. (2015). Learning Fast-Mixing Models for Structured Prediction. Proceedings of the 32nd International Conference on Machine Learning, in Proceedings of Machine Learning Research 37:1063-1072 Available from https://proceedings.mlr.press/v37/steinhardtb15.html.

Related Material