Data-Dependent Conditional Priors for Unsupervised Learning of Multimodal Data
One of the major shortcomings of variational autoencoders is the inability to produce generations from the individual modalities of data originating from mixture distributions. This is primarily due to the use of a simple isotropic Gaussian as the prior for the latent code in the ancestral sampling...
Main Authors: | , , |
---|---|
Format: | Article |
Language: | English |
Published: |
MDPI AG
2020-08-01
|
Series: | Entropy |
Subjects: | |
Online Access: | https://www.mdpi.com/1099-4300/22/8/888 |
_version_ | 1797558486144385024 |
---|---|
author | Frantzeska Lavda Magda Gregorová Alexandros Kalousis |
author_facet | Frantzeska Lavda Magda Gregorová Alexandros Kalousis |
author_sort | Frantzeska Lavda |
collection | DOAJ |
description | One of the major shortcomings of variational autoencoders is the inability to produce generations from the individual modalities of data originating from mixture distributions. This is primarily due to the use of a simple isotropic Gaussian as the prior for the latent code in the ancestral sampling procedure for data generations. In this paper, we propose a novel formulation of variational autoencoders, conditional prior VAE (CP-VAE), with a two-level generative process for the observed data where continuous <inline-formula><math display="inline"><semantics><mi mathvariant="bold">z</mi></semantics></math></inline-formula> and a discrete <inline-formula><math display="inline"><semantics><mi mathvariant="bold">c</mi></semantics></math></inline-formula> variables are introduced in addition to the observed variables <inline-formula><math display="inline"><semantics><mi mathvariant="bold">x</mi></semantics></math></inline-formula>. By learning data-dependent conditional priors, the new variational objective naturally encourages a better match between the posterior and prior conditionals, and the learning of the latent categories encoding the major source of variation of the original data in an unsupervised manner. Through sampling continuous latent code from the data-dependent conditional priors, we are able to generate new samples from the individual mixture components corresponding, to the multimodal structure over the original data. Moreover, we unify and analyse our objective under different independence assumptions for the joint distribution of the continuous and discrete latent variables. We provide an empirical evaluation on one synthetic dataset and three image datasets, FashionMNIST, MNIST, and Omniglot, illustrating the generative performance of our new model comparing to multiple baselines. |
first_indexed | 2024-03-10T17:32:05Z |
format | Article |
id | doaj.art-e1aaa28e2fba4819822de054f72380f5 |
institution | Directory Open Access Journal |
issn | 1099-4300 |
language | English |
last_indexed | 2024-03-10T17:32:05Z |
publishDate | 2020-08-01 |
publisher | MDPI AG |
record_format | Article |
series | Entropy |
spelling | doaj.art-e1aaa28e2fba4819822de054f72380f52023-11-20T10:00:36ZengMDPI AGEntropy1099-43002020-08-0122888810.3390/e22080888Data-Dependent Conditional Priors for Unsupervised Learning of Multimodal DataFrantzeska Lavda0Magda Gregorová1Alexandros Kalousis2Faculty of Science, Computer Science Department, University of Geneva, 1214 Geneva, SwitzerlandGeneva School of Business Administration (DMML Group), HES-SO, 1227 Geneva, SwitzerlandGeneva School of Business Administration (DMML Group), HES-SO, 1227 Geneva, SwitzerlandOne of the major shortcomings of variational autoencoders is the inability to produce generations from the individual modalities of data originating from mixture distributions. This is primarily due to the use of a simple isotropic Gaussian as the prior for the latent code in the ancestral sampling procedure for data generations. In this paper, we propose a novel formulation of variational autoencoders, conditional prior VAE (CP-VAE), with a two-level generative process for the observed data where continuous <inline-formula><math display="inline"><semantics><mi mathvariant="bold">z</mi></semantics></math></inline-formula> and a discrete <inline-formula><math display="inline"><semantics><mi mathvariant="bold">c</mi></semantics></math></inline-formula> variables are introduced in addition to the observed variables <inline-formula><math display="inline"><semantics><mi mathvariant="bold">x</mi></semantics></math></inline-formula>. By learning data-dependent conditional priors, the new variational objective naturally encourages a better match between the posterior and prior conditionals, and the learning of the latent categories encoding the major source of variation of the original data in an unsupervised manner. Through sampling continuous latent code from the data-dependent conditional priors, we are able to generate new samples from the individual mixture components corresponding, to the multimodal structure over the original data. Moreover, we unify and analyse our objective under different independence assumptions for the joint distribution of the continuous and discrete latent variables. We provide an empirical evaluation on one synthetic dataset and three image datasets, FashionMNIST, MNIST, and Omniglot, illustrating the generative performance of our new model comparing to multiple baselines.https://www.mdpi.com/1099-4300/22/8/888VAEgenerative modelslearned prior |
spellingShingle | Frantzeska Lavda Magda Gregorová Alexandros Kalousis Data-Dependent Conditional Priors for Unsupervised Learning of Multimodal Data Entropy VAE generative models learned prior |
title | Data-Dependent Conditional Priors for Unsupervised Learning of Multimodal Data |
title_full | Data-Dependent Conditional Priors for Unsupervised Learning of Multimodal Data |
title_fullStr | Data-Dependent Conditional Priors for Unsupervised Learning of Multimodal Data |
title_full_unstemmed | Data-Dependent Conditional Priors for Unsupervised Learning of Multimodal Data |
title_short | Data-Dependent Conditional Priors for Unsupervised Learning of Multimodal Data |
title_sort | data dependent conditional priors for unsupervised learning of multimodal data |
topic | VAE generative models learned prior |
url | https://www.mdpi.com/1099-4300/22/8/888 |
work_keys_str_mv | AT frantzeskalavda datadependentconditionalpriorsforunsupervisedlearningofmultimodaldata AT magdagregorova datadependentconditionalpriorsforunsupervisedlearningofmultimodaldata AT alexandroskalousis datadependentconditionalpriorsforunsupervisedlearningofmultimodaldata |