Mixture Density Hyperspherical Generative Adversarial Networks

Abstract

The Generative Adversarial Networks (GANs) are deep generative models that can generate realistic samples, but they are difficult to train in practice due to the problem of mode collapse, where the generator only repeatedly generates one mode in samples during the learning process, or only generates a small number of modes after reaching the Nash equilibrium during the adversarial training. In order to solve this issue while making the generator contains promising generation ability, we propose a mixture density hyperspherical generative model namely MDH-GAN that combines variational autoencoder (VAE) and generative adversarial network. Unlike most of the GAN-based generative models that consider a Gaussian prior, MDH-GAN adopts the von Mises-Fisher (vMF) prior defined on a unit hypersphere. Our model combines VAE with GAN by integrating the encoder of VAE with GAN to form a jointly training framework. Therefore, the generator of our model can learn data distribution with a hyperspherical latent structure, leading to an improved generative ability of the generator. Moreover, a vMF mixture model is deployed in the discriminator to form a hypersphere space to avoid mode collapse of the model. In our experiments, by calculating the Fréchet Inception distance (FID) between the generated images and real ones, we prove that MDH-GAN has a better ability to generate high-quality images with high diversity.