DS Ph.D. Qualifier Presentation | Kevin Hickey | Thurs. Nov. 21st @ 1:00pm | Commutative and Disentangled GAN Latent Space Navigation with Magnitude Control

DATA SCIENCE

Ph.D. Qualifier Presentation

Kevin Hickey

Thursday, November 21, 1:00pm

Location: Unity Hall 420

Committee: 

Dr. Elke Rundensteiner, Advisor, Computer Science, Data Science & Artificial Intelligence

Dr. Xiangnan Kong, Computer Science, Data Science

Dr. Oren Mangoubi, Mathematical Sciences, Data Science

Title: Commutative and Disentangled GAN Latent Space Navigation with Magnitude Control

Abstract: Controllable semantic image editing involves changing a particular characteristic of a given image, such as changing the degree of a smile in a portrait image. Many methods for achieving this use Generative Adversarial Networks (GANs), specifically manipulating the input latent variable such that a resulting newly generated image reflects the desired changes. Often, the desired changes should be disentangled, in that only one attribute changes at a time. However, most GAN disentanglement methods lack two key properties: 1.) commutativity, which states that edits of attributes in differing orders result in the same image, and 2.) explicit magnitude control, or the ability to edit an attribute by a desired amount. To mitigate these shortcomings, we propose CoMMGAN , a method that learns both properties simultaneously. We first provide empirical results on synthetic data, highlighting the shortcomings of related methods compared to our proposed approach, and investigate the performance of our method on real and synthetic images, providing quantitative and qualitative analysis of commutative and magnitude control properties.