Investigating the Learning Behavior of Generative Adversarial Networks

Abstract: Since their introduction in 2014, generative adversarial networks (GANs) have quickly become one of the most popular and successful frameworks for training deep generative models. GANs have shown exceptional results on different image generation tasks and they are known for their ability to produce realistic high- resolution images. However, despite the success and widespread use of the GAN framework, the models can be highly unstable to train and the training process has been shown to be extremely sensitive to hyperparameter settings and network architectures. In this thesis, we investigate how the latent dimension, batch size, and learning rate affect the stability and performance of both the non-saturating GAN (NS-GAN) and the Wasserstein GAN with added gradient penalty (WGAN-GP). Furthermore, we examine how the addition of noise to the inputs of the discriminator affects the stability of both GAN variants. The experiments are performed on three data sets – MNIST, CIFAR-10, and a grayscale version of CIFAR-10 – and all models are evaluated using both the Fréchet Inception Distance (FID) and precision and recall. The results presented in the thesis indicate that the learning rate has the largest impact on the stability and performance of both NS-GAN and WGANGP, and that the latent dimension and batch size have a relatively small impact when combined with an appropriate learning rate. Furthermore, we find that WGAN-GP outperforms NS-GAN on MNIST, while NS-GAN outperforms WGAN-GP on both versions of CIFAR-10. We also find that adding noise to the inputs of the discriminator during training greatly helps to stabilize the NS-GAN training process, while it has a limited effect on WGAN-GP.