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

Download:

Current browse context:

stat.ML

Change to browse by:

References & Citations

Bookmark

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

Statistics > Machine Learning

Title: Deep Generative Models with Stick-Breaking Priors

Abstract: Bayesian nonparametric models are attractive for their data-dependent capacity, but their implementation can be problematic due to computational or analytical obstacles. We make progress on this problem by extending Stochastic Gradient Variational Bayes (Kingma & Welling, 2013), a 'black box' method for approximate posterior inference, to stick-breaking priors (Ishwaran & James, 2001). This innovation allows us to define deep generative models (DGMs) with infinite dimensional latent variables. We experimentally demonstrate that DGMs with Dirichlet process priors learn highly discriminative latent representations that are well suited for semi-supervised settings and often outperform the popular Gaussian alternative.
Subjects: Machine Learning (stat.ML)
Cite as: arXiv:1605.06197 [stat.ML]
  (or arXiv:1605.06197v1 [stat.ML] for this version)

Submission history

From: Eric Nalisnick [view email]
[v1] Fri, 20 May 2016 02:20:13 GMT (2437kb,D)
[v2] Tue, 6 Dec 2016 03:16:25 GMT (2578kb,D)
[v3] Mon, 3 Apr 2017 22:36:33 GMT (2577kb,D)

Link back to: arXiv, form interface, contact.