Keywords: variational autoencoders, variational inference, data imputation, generative models
TL;DR: We propose a novel VAE-based framework learning from partially-observed data for imputation and generation.
Abstract: Despite promising progress on unimodal data imputation (e.g. image inpainting), models for multimodal data imputation are far from satisfactory. In this work, we propose variational selective autoencoder (VSAE) for this task. Learning only from partially-observed data, VSAE can model the joint distribution of observed/unobserved modalities and the imputation mask, resulting in a unified model for various down-stream tasks including data generation and imputation. Evaluation on synthetic high-dimensional and challenging low-dimensional multimodal datasets shows significant improvement over state-of-the-art imputation models.