Loading...
Representation learning to stabilize generative adversarial network with imbalanced, sparse, and irregular data
Seo, Hoon
Seo, Hoon
Citations
Altmetric:
Advisor
Editor
Date
Date Issued
2024
Date Submitted
Collections
Research Projects
Organizational Units
Journal Issue
Embargo Expires
2026-04-04
Abstract
Accurate analysis of data is essential across various fields, including geophysics and medical imaging. However, many datasets in these fields are often imbalanced, sparse, and irregular, complicating comprehensive analysis. This highlights the need for robust methods to handle these challenging data characteristics, including effective imputation and augmentation techniques. While imputation and data augmentation have been extensively studied in machine learning, many existing models assume a regular distribution of data and require a minimum density of observations. These constraints limit their effectiveness for the sparse and irregular patterns typical of many real-world datasets.
Given these challenges, deep learning approaches, particularly Generative Adversarial Networks (GANs), offer a promising alternative due to their ability to learn complex patterns from limited and irregular data. However, training GANs is often unstable, which presents a significant barrier to their broader application. This thesis explores the potential of representation learning to enhance the stability of GAN training for imbalanced, sparse, and irregular data. Specifically, it investigates how learning new representations of input data can improve the dynamics between the discriminator and generator in GANs, thereby enhancing performance.
The research has two main objectives: first, to identify the conditions under which GANs encounter instability issues when applied to various data imputation and augmentation tasks; and second, to examine how representation learning can mitigate these challenges and stabilize GAN training. This study aims to advance the field of data analysis by providing more robust and accurate tools for handling imbalanced, sparse, and irregular datasets, thereby supporting better decision-making and resource estimation efforts across diverse domains.
Associated Publications
Rights
Copyright of the original work is retained by the author.