@inproceedings{xu-durrett-2018-spherical,
title = "Spherical Latent Spaces for Stable Variational Autoencoders",
author = "Xu, Jiacheng and
Durrett, Greg",
editor = "Riloff, Ellen and
Chiang, David and
Hockenmaier, Julia and
Tsujii, Jun{'}ichi",
booktitle = "Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing",
month = oct # "-" # nov,
year = "2018",
address = "Brussels, Belgium",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/D18-1480",
doi = "10.18653/v1/D18-1480",
pages = "4503--4513",
abstract = "A hallmark of variational autoencoders (VAEs) for text processing is their combination of powerful encoder-decoder models, such as LSTMs, with simple latent distributions, typically multivariate Gaussians. These models pose a difficult optimization problem: there is an especially bad local optimum where the variational posterior always equals the prior and the model does not use the latent variable at all, a kind of {``}collapse{''} which is encouraged by the KL divergence term of the objective. In this work, we experiment with another choice of latent distribution, namely the von Mises-Fisher (vMF) distribution, which places mass on the surface of the unit hypersphere. With this choice of prior and posterior, the KL divergence term now only depends on the variance of the vMF distribution, giving us the ability to treat it as a fixed hyperparameter. We show that doing so not only averts the KL collapse, but consistently gives better likelihoods than Gaussians across a range of modeling conditions, including recurrent language modeling and bag-of-words document modeling. An analysis of the properties of our vMF representations shows that they learn richer and more nuanced structures in their latent representations than their Gaussian counterparts.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="xu-durrett-2018-spherical">
<titleInfo>
<title>Spherical Latent Spaces for Stable Variational Autoencoders</title>
</titleInfo>
<name type="personal">
<namePart type="given">Jiacheng</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Greg</namePart>
<namePart type="family">Durrett</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2018-oct-nov</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing</title>
</titleInfo>
<name type="personal">
<namePart type="given">Ellen</namePart>
<namePart type="family">Riloff</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">David</namePart>
<namePart type="family">Chiang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Julia</namePart>
<namePart type="family">Hockenmaier</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun’ichi</namePart>
<namePart type="family">Tsujii</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Brussels, Belgium</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>A hallmark of variational autoencoders (VAEs) for text processing is their combination of powerful encoder-decoder models, such as LSTMs, with simple latent distributions, typically multivariate Gaussians. These models pose a difficult optimization problem: there is an especially bad local optimum where the variational posterior always equals the prior and the model does not use the latent variable at all, a kind of “collapse” which is encouraged by the KL divergence term of the objective. In this work, we experiment with another choice of latent distribution, namely the von Mises-Fisher (vMF) distribution, which places mass on the surface of the unit hypersphere. With this choice of prior and posterior, the KL divergence term now only depends on the variance of the vMF distribution, giving us the ability to treat it as a fixed hyperparameter. We show that doing so not only averts the KL collapse, but consistently gives better likelihoods than Gaussians across a range of modeling conditions, including recurrent language modeling and bag-of-words document modeling. An analysis of the properties of our vMF representations shows that they learn richer and more nuanced structures in their latent representations than their Gaussian counterparts.</abstract>
<identifier type="citekey">xu-durrett-2018-spherical</identifier>
<identifier type="doi">10.18653/v1/D18-1480</identifier>
<location>
<url>https://aclanthology.org/D18-1480</url>
</location>
<part>
<date>2018-oct-nov</date>
<extent unit="page">
<start>4503</start>
<end>4513</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Spherical Latent Spaces for Stable Variational Autoencoders
%A Xu, Jiacheng
%A Durrett, Greg
%Y Riloff, Ellen
%Y Chiang, David
%Y Hockenmaier, Julia
%Y Tsujii, Jun’ichi
%S Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing
%D 2018
%8 oct nov
%I Association for Computational Linguistics
%C Brussels, Belgium
%F xu-durrett-2018-spherical
%X A hallmark of variational autoencoders (VAEs) for text processing is their combination of powerful encoder-decoder models, such as LSTMs, with simple latent distributions, typically multivariate Gaussians. These models pose a difficult optimization problem: there is an especially bad local optimum where the variational posterior always equals the prior and the model does not use the latent variable at all, a kind of “collapse” which is encouraged by the KL divergence term of the objective. In this work, we experiment with another choice of latent distribution, namely the von Mises-Fisher (vMF) distribution, which places mass on the surface of the unit hypersphere. With this choice of prior and posterior, the KL divergence term now only depends on the variance of the vMF distribution, giving us the ability to treat it as a fixed hyperparameter. We show that doing so not only averts the KL collapse, but consistently gives better likelihoods than Gaussians across a range of modeling conditions, including recurrent language modeling and bag-of-words document modeling. An analysis of the properties of our vMF representations shows that they learn richer and more nuanced structures in their latent representations than their Gaussian counterparts.
%R 10.18653/v1/D18-1480
%U https://aclanthology.org/D18-1480
%U https://doi.org/10.18653/v1/D18-1480
%P 4503-4513
Markdown (Informal)
[Spherical Latent Spaces for Stable Variational Autoencoders](https://aclanthology.org/D18-1480) (Xu & Durrett, EMNLP 2018)
ACL