Variational Generative Adversarial Networks for Preventing Mode Collapse

Abstract

Generative models try to obtain a probability distribution that is similar to that of observed data. Two different solutions have been proposed in this regard in recent years: one is to minimize the divergence (distance) between the two distributions by maximizing the variational lower bound, and the other is to implicitly reduce the distance between the two distributions through adversarial processes. One of the problems in generative adversarial networks (GANs) is the mode collapse. Mode collapse is a phenomenon in which, for various inputs, the generative model generates low variety or similar images. This paper tries to provide a solution to the mode collapse problem proposing a novel method called variational generative adversarial networks (VGANs). This method exploits variational autoencoders to initialize GANs. In other words, in addition to maximizing the variational lower bound, it also implicitly reduces the distance between the two distributions. Experimental results show that this method can deal with the mode collapse problem better than the state-of-the-art. Moreover, in the qualitative analysis, according to a survey of 136 people on the authenticity of the generated images, the proposed method can generate images more similar to real ones.

Keywords

• deep learning
• generative adversarial networks
• machine learning
• variational inference
• mode collapse