Adversarially Regularized Autoencoders

Junbo Zhao, Yoon Kim, Kelly Zhang, Alexander Rush, Yann LeCun
Proceedings of the 35th International Conference on Machine Learning, PMLR 80:5902-5911, 2018.

Abstract

Deep latent variable models, trained using variational autoencoders or generative adversarial networks, are now a key technique for representation learning of continuous structures. However, applying similar methods to discrete structures, such as text sequences or discretized images, has proven to be more challenging. In this work, we propose a more flexible method for training deep latent variable models of discrete structures. Our approach is based on the recently proposed Wasserstein Autoencoder (WAE) which formalizes adversarial autoencoders as an optimal transport problem. We first extend this framework to model discrete sequences, and then further explore different learned priors targeting a controllable representation. Unlike many other latent variable generative models for text, this adversarially regularized autoencoder (ARAE) allows us to generate fluent textual outputs as well as perform manipulations in the latent space to induce change in the output space. Finally we show that the latent representation can be trained to perform unaligned textual style transfer, giving improvements both in automatic measures and human evaluation.

Cite this Paper


BibTeX
@InProceedings{pmlr-v80-zhao18b, title = {Adversarially Regularized Autoencoders}, author = {Zhao, Junbo and Kim, Yoon and Zhang, Kelly and Rush, Alexander and LeCun, Yann}, booktitle = {Proceedings of the 35th International Conference on Machine Learning}, pages = {5902--5911}, year = {2018}, editor = {Dy, Jennifer and Krause, Andreas}, volume = {80}, series = {Proceedings of Machine Learning Research}, month = {10--15 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v80/zhao18b/zhao18b.pdf}, url = {http://proceedings.mlr.press/v80/zhao18b.html}, abstract = {Deep latent variable models, trained using variational autoencoders or generative adversarial networks, are now a key technique for representation learning of continuous structures. However, applying similar methods to discrete structures, such as text sequences or discretized images, has proven to be more challenging. In this work, we propose a more flexible method for training deep latent variable models of discrete structures. Our approach is based on the recently proposed Wasserstein Autoencoder (WAE) which formalizes adversarial autoencoders as an optimal transport problem. We first extend this framework to model discrete sequences, and then further explore different learned priors targeting a controllable representation. Unlike many other latent variable generative models for text, this adversarially regularized autoencoder (ARAE) allows us to generate fluent textual outputs as well as perform manipulations in the latent space to induce change in the output space. Finally we show that the latent representation can be trained to perform unaligned textual style transfer, giving improvements both in automatic measures and human evaluation.} }
Endnote
%0 Conference Paper %T Adversarially Regularized Autoencoders %A Junbo Zhao %A Yoon Kim %A Kelly Zhang %A Alexander Rush %A Yann LeCun %B Proceedings of the 35th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2018 %E Jennifer Dy %E Andreas Krause %F pmlr-v80-zhao18b %I PMLR %P 5902--5911 %U http://proceedings.mlr.press/v80/zhao18b.html %V 80 %X Deep latent variable models, trained using variational autoencoders or generative adversarial networks, are now a key technique for representation learning of continuous structures. However, applying similar methods to discrete structures, such as text sequences or discretized images, has proven to be more challenging. In this work, we propose a more flexible method for training deep latent variable models of discrete structures. Our approach is based on the recently proposed Wasserstein Autoencoder (WAE) which formalizes adversarial autoencoders as an optimal transport problem. We first extend this framework to model discrete sequences, and then further explore different learned priors targeting a controllable representation. Unlike many other latent variable generative models for text, this adversarially regularized autoencoder (ARAE) allows us to generate fluent textual outputs as well as perform manipulations in the latent space to induce change in the output space. Finally we show that the latent representation can be trained to perform unaligned textual style transfer, giving improvements both in automatic measures and human evaluation.
APA
Zhao, J., Kim, Y., Zhang, K., Rush, A. & LeCun, Y.. (2018). Adversarially Regularized Autoencoders. Proceedings of the 35th International Conference on Machine Learning, in Proceedings of Machine Learning Research 80:5902-5911 Available from http://proceedings.mlr.press/v80/zhao18b.html.

Related Material