Training Variational Autoencoders with Buffered Stochastic Variational Inference

Rui Shu, Hung Bui, Jay Whang, Stefano Ermon
Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics, PMLR 89:2134-2143, 2019.

Abstract

The recognition network in deep latent variable models such as variational autoencoders (VAEs) relies on amortized inference for efficient posterior approximation that can scale up to large datasets. However, this technique has also been demonstrated to select suboptimal variational parameters, often resulting in considerable additional error called the amortization gap. To close the amortization gap and improve the training of the generative model, recent works have introduced an additional refinement step that applies stochastic variational inference (SVI) to improve upon the variational parameters returned by the amortized inference model. In this paper, we propose the Buffered Stochastic Variational Inference (BSVI), a new refinement procedure that makes use of SVI’s sequence of intermediate variational proposal distributions and their corresponding importance weights to construct a new generalized importance-weighted lower bound. We demonstrate empirically that training the variational autoencoders with BSVI consistently out-performs SVI, yielding an improved training procedure for VAEs.

Cite this Paper


BibTeX
@InProceedings{pmlr-v89-shu19a, title = {Training Variational Autoencoders with Buffered Stochastic Variational Inference}, author = {Shu, Rui and Bui, Hung and Whang, Jay and Ermon, Stefano}, booktitle = {Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics}, pages = {2134--2143}, year = {2019}, editor = {Chaudhuri, Kamalika and Sugiyama, Masashi}, volume = {89}, series = {Proceedings of Machine Learning Research}, month = {16--18 Apr}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v89/shu19a/shu19a.pdf}, url = {https://proceedings.mlr.press/v89/shu19a.html}, abstract = {The recognition network in deep latent variable models such as variational autoencoders (VAEs) relies on amortized inference for efficient posterior approximation that can scale up to large datasets. However, this technique has also been demonstrated to select suboptimal variational parameters, often resulting in considerable additional error called the amortization gap. To close the amortization gap and improve the training of the generative model, recent works have introduced an additional refinement step that applies stochastic variational inference (SVI) to improve upon the variational parameters returned by the amortized inference model. In this paper, we propose the Buffered Stochastic Variational Inference (BSVI), a new refinement procedure that makes use of SVI’s sequence of intermediate variational proposal distributions and their corresponding importance weights to construct a new generalized importance-weighted lower bound. We demonstrate empirically that training the variational autoencoders with BSVI consistently out-performs SVI, yielding an improved training procedure for VAEs.} }
Endnote
%0 Conference Paper %T Training Variational Autoencoders with Buffered Stochastic Variational Inference %A Rui Shu %A Hung Bui %A Jay Whang %A Stefano Ermon %B Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics %C Proceedings of Machine Learning Research %D 2019 %E Kamalika Chaudhuri %E Masashi Sugiyama %F pmlr-v89-shu19a %I PMLR %P 2134--2143 %U https://proceedings.mlr.press/v89/shu19a.html %V 89 %X The recognition network in deep latent variable models such as variational autoencoders (VAEs) relies on amortized inference for efficient posterior approximation that can scale up to large datasets. However, this technique has also been demonstrated to select suboptimal variational parameters, often resulting in considerable additional error called the amortization gap. To close the amortization gap and improve the training of the generative model, recent works have introduced an additional refinement step that applies stochastic variational inference (SVI) to improve upon the variational parameters returned by the amortized inference model. In this paper, we propose the Buffered Stochastic Variational Inference (BSVI), a new refinement procedure that makes use of SVI’s sequence of intermediate variational proposal distributions and their corresponding importance weights to construct a new generalized importance-weighted lower bound. We demonstrate empirically that training the variational autoencoders with BSVI consistently out-performs SVI, yielding an improved training procedure for VAEs.
APA
Shu, R., Bui, H., Whang, J. & Ermon, S.. (2019). Training Variational Autoencoders with Buffered Stochastic Variational Inference. Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics, in Proceedings of Machine Learning Research 89:2134-2143 Available from https://proceedings.mlr.press/v89/shu19a.html.

Related Material