Adversarial Variational Bayes: Unifying Variational Autoencoders and Generative Adversarial Networks

• DOI: 10.48550/arxiv.1701.04722
• Publication Date: 2017-01-01
• Authors: Lars Mescheder; Sebastian Nowozin; Andreas Geiger
• Keywords: Discriminative model; Generative model

Variational Autoencoders (VAEs) are expressive latent variable models that can be used to learn complex probability distributions from training data. However, the quality of resulting model crucially relies on expressiveness inference model. We introduce Adversarial Bayes (AVB), a technique for with arbitrarily models. achieve this by introducing an auxiliary discriminative network allows rephrase maximum-likelihood-problem as two-player game, hence establishing principled connection between VAEs and Generative Networks (GANs). show in nonparametric limit our method yields exact maximum-likelihood assignment parameters generative model, well posterior distribution over variables given observation. Contrary competing approaches which combine GANs, approach has clear theoretical justification, retains most advantages standard is easy implement.