A new form of variational autoencoder (VAE) is developed, in which the joint distribution of data and codes is considered in two (symmetric) forms: (i) from observed data fed through the encoder to yield codes, and (ii) from latent codes drawn from a simple prior and propagated through the decoder to manifest data. Lower bounds are learned for marginal log-likelihood fits observed data and latent codes. When learning with the variational bound, one seeks to minimize the symmetric Kullback-Leibler divergence of joint density functions from (i) and (ii), while simultaneously seeking to maximize the two marginal log-likelihoods. To facilitate learning, a new form of adversarial training is developed. An extensive set of experiments is performed, in which we demonstrate state-of-the-art data reconstruction and generation on several image benchmarks datasets.

The paper proposes a variant of the Variational Auto-Encoder training objective. It uses adversarial training, to minimize a symmetric KL divergence between the joint distributions of latent and observed variables p(z,x) p(z)p_\theta(x z) and q(z,x) q(x)q_\phi(z x) . The approach is similar to the recent [ Mescheder, Nowozin, Geiger. Adversarial variational bayes: Unifying variational autoencoders and generative adversarial networks, 2016 ] in its joining VAE and GAN-like objective, but it is original in that it minimizes a symmetric KL divergence (with a GAN-like objective), which appears crucial to achieve good quality samples. It is also reminiscent of ALI [ Dumoulin et al.

A new form of variational autoencoder (VAE) is developed, in which the joint distribution of data and codes is considered in two (symmetric) forms: (i) from observed data fed through the encoder to yield codes, and (ii) from latent codes drawn from a simple prior and propagated through the decoder to manifest data. Lower bounds are learned for marginal log-likelihood fits observed data and latent codes. When learning with the variational bound, one seeks to minimize the symmetric Kullback-Leibler divergence of joint density functions from (i) and (ii), while simultaneously seeking to maximize the two marginal log-likelihoods. To facilitate learning, a new form of adversarial training is developed. An extensive set of experiments is performed, in which we demonstrate state-of-the-art data reconstruction and generation on several image benchmarks datasets.