Random Generative Adversarial Networks

Random Generative Adversarial Networks

연구 분야: Artificial Intelligence

학회: SoICT '22: Proceedings of the 11th International Symposium on Information and Communication Technology

Generative Adversarial Networks have surprisingly shown great ability in synthesizing high-fidelity and diverse images while resolving the problem of so-called mode collapse still remains a challenge to all researches in this field. In this paper, we propose a new GANs method called Random Generative Adversarial Networks (RandomGANs), a combination of Random Forest algorithm and Generative Adversarial Networks, in which utilizes multiple discriminators where each discriminator is considered as a decision tree and each of them will be separately trained on a separated dataset that sampled without replacement from the original dataset. We empirically demonstrate (1) the quality of generated images in RandomGANs and their strong similarity to real images, (2) show how different knowledge each discriminator learns by comparing with others and overcoming the mode-collapse problem by encouraging the Discriminator to discover a wide range of different modes in the input data, and (3) indicate how our model betters in stabilizing the training process between generator and discriminator part.