Improving Discriminator-Generator Balance in Generative Adversarial Networks

Simen Selseng and Björn Gambäck

Feb 15, 2018 (modified: Feb 15, 2018) ICLR 2018 Conference Blind Submission readers: everyone Show Bibtex
  • Abstract: Training Artificial Neural Networks to generate better images is a hard problem to solve with supervised techniques. Generative Adversarial Networks (GAN) is an unsupervised approach that utilizes two ANNs in one system. The first neural network is called the Discriminator and evaluates the quality of images generated by the other network, the Generator. However, image generation has no fixed solution, making evaluation difficult. This can be addressed through comparing two GAN models by letting them evaluate each other; the paper proposes a custom implementation of this evaluation method that allows for comparison between two GAN models both overall and throughout the learning process. GANs have a reputation for being hard to train. One concrete problem is main- taining the balance between the Generator and Discriminator. As for humans, it is easier to rate the quality of images then it is to actually create them. A good evaluator is necessary, but it must not out-power the generative model. The paper explores two main approaches to achieve a better balanced GAN model. The first method makes guided alterations to the usually random input of the Generator. The second method adds an additional Discriminator to the model. Techniques based on both of these methods were shown to effectively guide the training pro- cess and creating strong models that outperformed regular GANs when compared using the previously mentioned evaluation method.