Amortized Structured Stochastic Variational Inference for Gaussian Process Latent Variable Models
Paper applies amortized structured stochastic variational inference to Gaussian Process Latent Variable Models, letting latent posteriors depend on inducing points and improving manifold reconstruction metrics.
The paper addresses Gaussian Process Latent Variable Models (GPLVMs), where mean-field variational approximations between GP inducing points and latent variables limit uncertainty estimation of the learned manifold. The authors apply Amortized Structured Stochastic Variational Inference so the variational posterior over latent space is conditionally dependent on inducing point values. Experiments demonstrate the more flexible posterior improves several reconstruction metrics on the data manifold.
- Structured variational posterior makes latent variables conditionally dependent on GP inducing points
- Improves reconstruction metrics for learned low-dimensional data manifolds
- Addresses epistemic uncertainty limits of mean-field approximations in GPLVMs
From the original article
1 min read at arxiv.orgMany machine learning methods aim to approximate the lower-dimensional manifold on which the data lives. A desirable feature of such methods is that they should capture the epistemic uncertainty of this learned manifold. One model that achieves this is the Gaussian Process Latent Variable Model, in which a Gaussian Process (GP) mapping from the latent space provides an estimate of the uncertainty of the manifold.
By Maksym Tretiakov · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.