missIWAE: Deep Generative Modelling and Imputation of Incomplete Data

12/06/2018

∙

by   Pierre-Alexandre Mattei, et al.

∙

0

∙

share

We present a simple technique to train deep latent variable models (DLVMs) when the training set contains missing data. Our approach is based on the importance-weighted autoencoder (IWAE) of Burda et al. (2016), and also allows single or multiple imputation of the incomplete data set. We illustrate it by training a convolutional DLVM on a static binarisation of MNIST that contains 50 network that classifies these incomplete digits as well as complete ones.