Revisiting Reweighted Wake-Sleep for Models with Stochastic Control Flow

Tuan Anh Le, Adam R. Kosiorek, N. Siddharth, Yee Whye Teh, Frank Wood
Proceedings of The 35th Uncertainty in Artificial Intelligence Conference, PMLR 115:1039-1049, 2020.

Abstract

Stochastic control-flow models (SCFMs) are a class of generative models that involve branching on choices from discrete random variables. Amortized gradient-based learning of SCFMs is challenging as most approaches targeting discrete variables rely on their continuous relaxations—which can be intractable in SCFMs, as branching on relaxations requires evaluating all (exponentially many) branching paths. Tractable alternatives mainly combine REINFORCE with complex control-variate schemes to improve the variance of naive estimators. Here, we revisit the reweighted wake-sleep (RWS) [5] algorithm, and through extensive evaluations, show that it outperforms current state-of-the-art methods in learning SCFMs. Further, in contrast to the importance weighted autoencoder, we observe that RWS learns better models and inference networks with increasing numbers of particles. Our results suggest that RWS is a competitive, often preferable, alternative for learning SCFMs.

Cite this Paper

BibTeX
@InProceedings{pmlr-v115-le20a, title = {Revisiting Reweighted Wake-Sleep for Models with Stochastic Control Flow}, author = {Le, Tuan Anh and Kosiorek, Adam R. and Siddharth, N. and Teh, Yee Whye and Wood, Frank}, booktitle = {Proceedings of The 35th Uncertainty in Artificial Intelligence Conference}, pages = {1039--1049}, year = {2020}, editor = {Adams, Ryan P. and Gogate, Vibhav}, volume = {115}, series = {Proceedings of Machine Learning Research}, month = {22--25 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v115/le20a/le20a.pdf}, url = {https://proceedings.mlr.press/v115/le20a.html}, abstract = {Stochastic control-flow models (SCFMs) are a class of generative models that involve branching on choices from discrete random variables. Amortized gradient-based learning of SCFMs is challenging as most approaches targeting discrete variables rely on their continuous relaxations—which can be intractable in SCFMs, as branching on relaxations requires evaluating all (exponentially many) branching paths. Tractable alternatives mainly combine REINFORCE with complex control-variate schemes to improve the variance of naive estimators. Here, we revisit the reweighted wake-sleep (RWS) [5] algorithm, and through extensive evaluations, show that it outperforms current state-of-the-art methods in learning SCFMs. Further, in contrast to the importance weighted autoencoder, we observe that RWS learns better models and inference networks with increasing numbers of particles. Our results suggest that RWS is a competitive, often preferable, alternative for learning SCFMs.} }
Endnote
%0 Conference Paper %T Revisiting Reweighted Wake-Sleep for Models with Stochastic Control Flow %A Tuan Anh Le %A Adam R. Kosiorek %A N. Siddharth %A Yee Whye Teh %A Frank Wood %B Proceedings of The 35th Uncertainty in Artificial Intelligence Conference %C Proceedings of Machine Learning Research %D 2020 %E Ryan P. Adams %E Vibhav Gogate %F pmlr-v115-le20a %I PMLR %P 1039--1049 %U https://proceedings.mlr.press/v115/le20a.html %V 115 %X Stochastic control-flow models (SCFMs) are a class of generative models that involve branching on choices from discrete random variables. Amortized gradient-based learning of SCFMs is challenging as most approaches targeting discrete variables rely on their continuous relaxations—which can be intractable in SCFMs, as branching on relaxations requires evaluating all (exponentially many) branching paths. Tractable alternatives mainly combine REINFORCE with complex control-variate schemes to improve the variance of naive estimators. Here, we revisit the reweighted wake-sleep (RWS) [5] algorithm, and through extensive evaluations, show that it outperforms current state-of-the-art methods in learning SCFMs. Further, in contrast to the importance weighted autoencoder, we observe that RWS learns better models and inference networks with increasing numbers of particles. Our results suggest that RWS is a competitive, often preferable, alternative for learning SCFMs.
APA
Le, T.A., Kosiorek, A.R., Siddharth, N., Teh, Y.W. & Wood, F.. (2020). Revisiting Reweighted Wake-Sleep for Models with Stochastic Control Flow. Proceedings of The 35th Uncertainty in Artificial Intelligence Conference, in Proceedings of Machine Learning Research 115:1039-1049 Available from https://proceedings.mlr.press/v115/le20a.html.

Related Material