References & Citations
Statistics > Machine Learning
Title: Deep Generative Models with Stick-Breaking Priors
(Submitted on 20 May 2016 (this version), latest version 3 Apr 2017 (v3))
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.
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.