Time: Monday, April 03, 2017, 1:00pm @ MTH3206

Speaker: Addison Bohannon (UMD)

Title: Variational inference with deep generative models

Abstract: The variational autoencoder framework proposed in Kingma and Welling, 2013 and Rezende, et al., 2014 has inspired numerous applications and extensions (e.g. Larsen, et al., 2015, Burda, et al., 2015, Xu, et al., 2015). This formulation has enabled the application of computational tools from deep learning to variational Bayesian inference and facilitated the unsupervised discovery of latent structure in generative models. We review the derivation of the variational autoencoder objective function with particular focus on the “re-parameterization trick.”

Back to seminar