Master's Thesis Defense: BayesBiGAN and GANify - Research and a Developer Tool for Generative Adversarial Networks

June 6, 2019
1:00 PM
Halligan 209
Speaker: Daniel Dinjian
Host: Liping Liu

Abstract

This thesis addresses the current state of Generative Adversarial Networks (GANs) in the field of Machine Learning, both to further the scientific, and non-scientific communities’ knowledge of the field. The first part of the work develops a new learning model, Bayesian Bidirectional GAN, that combines the BayesGAN and the BiGAN, dubbed the BayesBiGAN. The BayesBiGAN inherits the increased scope of functionality of the BiGAN that comes from training an encoder model to convert data back to latent space. Additionally, it extends the BiGAN by learning posterior distributions using the stochastic gradient Hamiltonian Monte Carlo method of the BayesGAN rather than conventional point-estimate networks. Converting to and from the latent space with posterior distributions will allow us to capture multimodality in our dataset and help us to better retain whatever underlying feature-representations the dataset has. The second part of the work addresses the issue of the disconnect between research and actual product development, especially pertaining to GANs. There has recently been a great deal of scientific literature on the creation and applications of GANs, yet there are very few actual applications in circulation. Research seems focused solely on demonstrating what can be done, but this information is published in research papers and often fails to make the next step into a useful product. In this part of the work, I create an API that will allow developers to use CycleGAN for style transfer to convert photos into painting renderings in the style of various famous painters. I also use this API to create my own educational webapp that allows people who don’t research machine learning to still use and understand this innovative GAN application.