In this paper, we propose an arbitrarily-conditioned data imputation framework built upon variational autoencoders and normalizing flows. The proposed model is capable of mapping any partial data to a multi-modal latent variational distribution. Sampling from such a distribution leads to stochastic imputation. Preliminary evaluation on MNIST dataset shows promising stochastic imputation conditioned on partial images as input.


title ={{Arbitrarily-conditioned Data Imputation}},
author ={Carvalho, Micael and Durand, Thibaut and He, Jiawei and Mehrasa, Nazanin and Mori, Greg},
booktitle ={Proceedings of The 2nd Symposium on Advances in Approximate Bayesian Inference},
year ={2019},
url = {},