Stick-Breaking Variational Autoencoders

Eric Nalisnick, Padhraic Smyth

Nov 04, 2016 (modified: Apr 03, 2017) ICLR 2017 conference submission readers: everyone
  • Abstract: We extend Stochastic Gradient Variational Bayes to perform posterior inference for the weights of Stick-Breaking processes. This development allows us to define a Stick-Breaking Variational Autoencoder (SB-VAE), a Bayesian nonparametric version of the variational autoencoder that has a latent representation with stochastic dimensionality. We experimentally demonstrate that the SB-VAE, and a semi-supervised variant, learn highly discriminative latent representations that often outperform the Gaussian VAE’s.
  • TL;DR: We define a variational autoencoder variant with stick-breaking latent variables thereby giving it adaptive width.
  • Keywords: Deep learning, Unsupervised Learning, Semi-Supervised Learning
  • Conflicts:,,