Exploring Normalizing Flows and Their Diverse Applications

Abstract

Generative adversarial networks (GANs) have seen widespread use in computer graphics and computer vision over recent years. However, they come with several drawbacks, including mode collapse, unstable training, and ambiguous convergence criteria. Moreover, conditional GANs often generate samples with limited diversity, posing significant challenges for many applications. In this talk, I will introduce normalizing flows, a unique type of generative network, discussing both their advantages and limitations. Following this, I will present various problems we have addressed using normalizing flows, such as icon colorization, basketball play synthesis, shuttle landing distribution modeling, noisy label training, and substituting datasets with normalizing flows.

Bio

Yu-Shuen Wang (王昱舜) is a professor of the Department of Computer Science at National Yang Ming Chiao Tung University. He received his PhD degree from the Visual System Laboratory, National Cheng Kung University, Tainan, Taiwan, ROC, in 2010. Currently, he lead the Computer Graphics and Visualization Lab at the Institute of Multimedia Engineering. His research interests include Computer Graphics, Computer Vision, Data Visualization, and Machine Learning. He was honored with the prestigious Wu Da-Yu Memorial Award and the NCTU EECS Outstanding Young Scholar Award in 2016.