Accurate and conservative estimates of MRF log-likelihood using reverse annealing

Yuri Burda, Roger Grosse, Ruslan Salakhutdinov
Proceedings of the Eighteenth International Conference on Artificial Intelligence and Statistics, PMLR 38:102-110, 2015.

Abstract

Markov random fields (MRFs) are difficult to evaluate as generative models because computing the test log-probabilities requires the intractable partition function. Annealed importance sampling (AIS) is widely used to estimate MRF partition functions, and often yields quite accurate results. However, AIS is prone to overestimate the log-likelihood with little indication that anything is wrong. We present the Reverse AIS Estimator (RAISE), a stochastic lower bound on the log-likelihood of an approximation to the original MRF model. RAISE requires only the same MCMC transition operators as standard AIS. Experimental results indicate that RAISE agrees closely with AIS log-probability estimates for RBMs, DBMs, and DBNs, but typically errs on the side of underestimating, rather than overestimating, the log-likelihood.

Cite this Paper

BibTeX
@InProceedings{pmlr-v38-burda15, title = {{Accurate and conservative estimates of MRF log-likelihood using reverse annealing}}, author = {Burda, Yuri and Grosse, Roger and Salakhutdinov, Ruslan}, booktitle = {Proceedings of the Eighteenth International Conference on Artificial Intelligence and Statistics}, pages = {102--110}, year = {2015}, editor = {Lebanon, Guy and Vishwanathan, S. V. N.}, volume = {38}, series = {Proceedings of Machine Learning Research}, address = {San Diego, California, USA}, month = {09--12 May}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v38/burda15.pdf}, url = {https://proceedings.mlr.press/v38/burda15.html}, abstract = {Markov random fields (MRFs) are difficult to evaluate as generative models because computing the test log-probabilities requires the intractable partition function. Annealed importance sampling (AIS) is widely used to estimate MRF partition functions, and often yields quite accurate results. However, AIS is prone to overestimate the log-likelihood with little indication that anything is wrong. We present the Reverse AIS Estimator (RAISE), a stochastic lower bound on the log-likelihood of an approximation to the original MRF model. RAISE requires only the same MCMC transition operators as standard AIS. Experimental results indicate that RAISE agrees closely with AIS log-probability estimates for RBMs, DBMs, and DBNs, but typically errs on the side of underestimating, rather than overestimating, the log-likelihood.} }
Endnote
%0 Conference Paper %T Accurate and conservative estimates of MRF log-likelihood using reverse annealing %A Yuri Burda %A Roger Grosse %A Ruslan Salakhutdinov %B Proceedings of the Eighteenth International Conference on Artificial Intelligence and Statistics %C Proceedings of Machine Learning Research %D 2015 %E Guy Lebanon %E S. V. N. Vishwanathan %F pmlr-v38-burda15 %I PMLR %P 102--110 %U https://proceedings.mlr.press/v38/burda15.html %V 38 %X Markov random fields (MRFs) are difficult to evaluate as generative models because computing the test log-probabilities requires the intractable partition function. Annealed importance sampling (AIS) is widely used to estimate MRF partition functions, and often yields quite accurate results. However, AIS is prone to overestimate the log-likelihood with little indication that anything is wrong. We present the Reverse AIS Estimator (RAISE), a stochastic lower bound on the log-likelihood of an approximation to the original MRF model. RAISE requires only the same MCMC transition operators as standard AIS. Experimental results indicate that RAISE agrees closely with AIS log-probability estimates for RBMs, DBMs, and DBNs, but typically errs on the side of underestimating, rather than overestimating, the log-likelihood.
RIS
TY - CPAPER TI - Accurate and conservative estimates of MRF log-likelihood using reverse annealing AU - Yuri Burda AU - Roger Grosse AU - Ruslan Salakhutdinov BT - Proceedings of the Eighteenth International Conference on Artificial Intelligence and Statistics DA - 2015/02/21 ED - Guy Lebanon ED - S. V. N. Vishwanathan ID - pmlr-v38-burda15 PB - PMLR DP - Proceedings of Machine Learning Research VL - 38 SP - 102 EP - 110 L1 - http://proceedings.mlr.press/v38/burda15.pdf UR - https://proceedings.mlr.press/v38/burda15.html AB - Markov random fields (MRFs) are difficult to evaluate as generative models because computing the test log-probabilities requires the intractable partition function. Annealed importance sampling (AIS) is widely used to estimate MRF partition functions, and often yields quite accurate results. However, AIS is prone to overestimate the log-likelihood with little indication that anything is wrong. We present the Reverse AIS Estimator (RAISE), a stochastic lower bound on the log-likelihood of an approximation to the original MRF model. RAISE requires only the same MCMC transition operators as standard AIS. Experimental results indicate that RAISE agrees closely with AIS log-probability estimates for RBMs, DBMs, and DBNs, but typically errs on the side of underestimating, rather than overestimating, the log-likelihood. ER -
APA
Burda, Y., Grosse, R. & Salakhutdinov, R.. (2015). Accurate and conservative estimates of MRF log-likelihood using reverse annealing. Proceedings of the Eighteenth International Conference on Artificial Intelligence and Statistics, in Proceedings of Machine Learning Research 38:102-110 Available from https://proceedings.mlr.press/v38/burda15.html.

Related Material