GAN model development
In this project we propose the Deli-Fisher GAN, a Generative Adversarial Network (GAN) that generates photo-realistic images by translating random data into meaningful data distributions.
In our framework, instead of using a single uniform or Gaussian distribution as source of random data,
we use mixture probability distributions to model the latent space and in addition, we include training on the distribution parameters. Furthermore, to improve stability and efficiency, we use the Fisher Integral Probability Metric as the divergence measure in our GAN model, instead of the usual Jensen Shannon divergence. We also show by experiments that our GAN performs no worse than the Fisher GAN, and that our GAN sometimes out-competes the Fisher GAN in terms of inception score while generating images with interesting properties.