Dual Importance Weight GAN

Gahye Lee, Seungkyu Lee

Abstract: Generative Adversarial Networks (GAN) are trained to generate a sample image of interest. To this end, generative network of GAN learns implicit distribution of true dataset from the classification samples with candidate generated samples. However, in real implementation of GAN, training the generative network with limited number of candidate samples guarantees to properly represent neither true distribution nor the distribution of generator outputs. In this paper, we propose dual importance weights for the candidate samples represented in the latent space of auto-encoder. The auto-encoder is pre-trained with real target dataset. Therefore, the latent space representation allows us to compare real distribution and the distribution of generated samples explicitly. Dual importance weights iteratively maximize the representation of generated samples for both distributions: current generator outputs and real dataset. Proposed generative model not only resolves mode collapse problem of GAN but also improves the convergence on target distribution. Experimental evaluation shows that the proposed network learns complete modes of target distribution more stable and faster than state of the art methods.