Gaussian Process Latent Variable Flows for Massively Missing Data

Vidhi Lalchand; Aditya Ravuri; Neil D. Lawrence

Back to publications

Gaussian Process Latent Variable Flows for Massively Missing Data

Vidhi Lalchand, Aditya Ravuri, Neil D. Lawrence

Third Symposium on Advances in Approximate Bayesian Inference, 2021.

Abstract

Gaussian process latent variable models (GPLVM) are used to perform nonlinear and probabilistic dimensionality reduction. They extend Gaussian processes (GP) to the domain of unsupervised learning. The Bayesian incarnation of the GPLVM uses a variational framework, where the posterior over all unknown quantities is approximated by a well-behaved variational family, a factorised Gaussian. This gives not only implicit regularisation but also mathematical convenience. In this work we narrow our focus on examining the quality of the latent representation learnt under this Gaussian assumption. We introduce non-Gaussianity in the distribution of the latent space through normalising flows. The flexibility afforded by flows is critical in modelling massively missing data. Inference is performed using Stochastic Variational Inference (SVI) with a structured variational lower bound that factorizes across data points permitting efficient and scalable mini-batching of gradients. We call this flexible model class (GPLVF). We compare this framework with traditional models like the Bayesian GPLVM. Our experiments focus on massively missing data settings.

Links

Cite this Paper

BibTeX


@InProceedings{publications/gaussian-process-latent-variable-flows-for-massively-missing-data,
  title = 	 {Gaussian Process Latent Variable Flows for Massively Missing Data},
  author = 	 {Lalchand, Vidhi and Ravuri, Aditya and Lawrence, Neil D.},
  booktitle = 	 {Third Symposium on Advances in Approximate Bayesian Inference},
  year = 	 {2021},
  url = 	 {/publications/gaussian-process-latent-variable-flows-for-massively-missing-data.html},
  abstract = 	 {Gaussian process latent variable models (GPLVM) are used to perform nonlinear and probabilistic dimensionality reduction. They extend Gaussian processes (GP) to the domain of unsupervised learning. The Bayesian incarnation of the GPLVM uses a variational framework, where the posterior over all unknown quantities is approximated by a well-behaved variational family, a factorised Gaussian. This gives not only implicit regularisation but also mathematical convenience. In this work we narrow our focus on examining the quality of the latent representation learnt under this Gaussian assumption. We introduce non-Gaussianity in the distribution of the latent space through normalising flows. The flexibility afforded by flows is critical in modelling massively missing data. Inference is performed using Stochastic Variational Inference (SVI) with a structured variational lower bound that factorizes across data points permitting efficient and scalable mini-batching of gradients. We call this flexible model class (GPLVF). We compare this framework with traditional models like the Bayesian GPLVM. Our experiments focus on massively missing data settings.}
}

Endnote

%0 Conference Paper
%T Gaussian Process Latent Variable Flows for Massively Missing Data
%A Vidhi Lalchand
%A Aditya Ravuri
%A Neil D. Lawrence
%B Third Symposium on Advances in Approximate Bayesian Inference
%D 2021	
%F publications/gaussian-process-latent-variable-flows-for-massively-missing-data
%U /publications/gaussian-process-latent-variable-flows-for-massively-missing-data.html
%X Gaussian process latent variable models (GPLVM) are used to perform nonlinear and probabilistic dimensionality reduction. They extend Gaussian processes (GP) to the domain of unsupervised learning. The Bayesian incarnation of the GPLVM uses a variational framework, where the posterior over all unknown quantities is approximated by a well-behaved variational family, a factorised Gaussian. This gives not only implicit regularisation but also mathematical convenience. In this work we narrow our focus on examining the quality of the latent representation learnt under this Gaussian assumption. We introduce non-Gaussianity in the distribution of the latent space through normalising flows. The flexibility afforded by flows is critical in modelling massively missing data. Inference is performed using Stochastic Variational Inference (SVI) with a structured variational lower bound that factorizes across data points permitting efficient and scalable mini-batching of gradients. We call this flexible model class (GPLVF). We compare this framework with traditional models like the Bayesian GPLVM. Our experiments focus on massively missing data settings.

RIS


TY  - CPAPER
TI  - Gaussian Process Latent Variable Flows for Massively Missing Data
AU  - Vidhi Lalchand
AU  - Aditya Ravuri
AU  - Neil D. Lawrence
BT  - Third Symposium on Advances in Approximate Bayesian Inference
DA  - 2021/01/15	
ID  - publications/gaussian-process-latent-variable-flows-for-massively-missing-data
UR  - /publications/gaussian-process-latent-variable-flows-for-massively-missing-data.html
AB  - Gaussian process latent variable models (GPLVM) are used to perform nonlinear and probabilistic dimensionality reduction. They extend Gaussian processes (GP) to the domain of unsupervised learning. The Bayesian incarnation of the GPLVM uses a variational framework, where the posterior over all unknown quantities is approximated by a well-behaved variational family, a factorised Gaussian. This gives not only implicit regularisation but also mathematical convenience. In this work we narrow our focus on examining the quality of the latent representation learnt under this Gaussian assumption. We introduce non-Gaussianity in the distribution of the latent space through normalising flows. The flexibility afforded by flows is critical in modelling massively missing data. Inference is performed using Stochastic Variational Inference (SVI) with a structured variational lower bound that factorizes across data points permitting efficient and scalable mini-batching of gradients. We call this flexible model class (GPLVF). We compare this framework with traditional models like the Bayesian GPLVM. Our experiments focus on massively missing data settings.
ER  -

APA


Lalchand, V., Ravuri, A. & Lawrence, N.D.. (2021). Gaussian Process Latent Variable Flows for Massively Missing Data. Third Symposium on Advances in Approximate Bayesian Inference Available from /publications/gaussian-process-latent-variable-flows-for-massively-missing-data.html.