GAN with autoencoder and importance sampling

Deep generative model such as generative adversarial networks (GAN) has shown impressive achievements in computer graphics applications. GAN is trained to learn the distribution of target data and is able to generate new samples similar to the original target data. However, most GAN based networks encounter mode collapse problem resulting in the generation of samples only from a single or a few modes of target data distribution. In order to address mode collapse problem, we propose to adopt autoencoder to learn target data distribution in encoded space. An importance sampling scheme is used to collect fake and real data samples in the encoded latent space and calculate the similarity of two distributions in real data space. Experimental evaluation compared to state-of-the-art method on synthetic and MNIST datasets shows the potential of our approach in reducing mode collapse problem and generating samples from diverse aspect of target data.