We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.LG

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo

Computer Science > Machine Learning

Title: Variational Autoencoders with Riemannian Brownian Motion Priors

Abstract: Variational Autoencoders (VAEs) represent the given data in a low-dimensional latent space, which is generally assumed to be Euclidean. This assumption naturally leads to the common choice of a standard Gaussian prior over continuous latent variables. Recent work has, however, shown that this prior has a detrimental effect on model capacity, leading to subpar performance. We propose that the Euclidean assumption lies at the heart of this failure mode. To counter this, we assume a Riemannian structure over the latent space, which constitutes a more principled geometric view of the latent codes, and replace the standard Gaussian prior with a Riemannian Brownian motion prior. We propose an efficient inference scheme that does not rely on the unknown normalizing factor of this prior. Finally, we demonstrate that this prior significantly increases model capacity using only one additional scalar parameter.
Comments: Published in ICML 2020
Subjects: Machine Learning (cs.LG); Machine Learning (stat.ML)
Journal reference: Proceedings of the 37th International Conference on Machine Learning, Vienna, Austria, PMLR 119, 2020
Cite as: arXiv:2002.05227 [cs.LG]
  (or arXiv:2002.05227v3 [cs.LG] for this version)

Submission history

From: Dimitris Kalatzis [view email]
[v1] Wed, 12 Feb 2020 20:35:21 GMT (2965kb,D)
[v2] Tue, 18 Feb 2020 21:08:37 GMT (2965kb,D)
[v3] Fri, 7 Aug 2020 16:01:28 GMT (7669kb,D)

Link back to: arXiv, form interface, contact.