Inverting VAEs for Improved Generative Accuracy

Recent advances in semi-supervised learning with deep generative models have shown promise in generalizing from small labeled datasets (xl,yl) to large unlabeled ones (xu). When the codomain (y) has known structure, a large unfeatured dataset (yu) is potentially available. We develop a parameter-efficient, deep semisupervised generative model for the purpose of exploiting this untapped data source. Empirical results show improved performance in disentangling variable semantics as well as improved discriminative prediction on a new MNIST task.