A Latent Space Understandable Generative Adversarial Network: SelfExGAN

This paper describes a latent space understandable network: Self- excited Generative Adversarial Network (Self- ExGAN), a novel self- excited structure based on adversarial learning. Compared with the conventional generative adversarial networks, SelfExGAN consists of three components, which are en- coder (E), generator (G), and discriminator (D). Different from other papers which directly apply reconstruction loss between encoder and generator, SelfExGAN introduces Nash equilibrium between these three parts in order to discover the correspondence between latent inputs and training data spontaneously. The most attractive point of SelfExGAN is that it can use the learned correspondence to guide G to generate homomorphic samples given different latent inputs in an unsupervised learning manner. We demonstrate the learning ability of SelfExGAN on MNIST, celebA, and SVHN datasets. Our code is available on https://github.com/LiamLYJ/SelfexGAN.