Machine learning approaches to star-galaxy classification

Accurate star-galaxy classification has many important applications in modern precision cosmology. However, a vast number of faint sources that are detected in the current and next-generation ground-based surveys may be challenged by poor star-galaxy classification. Thus, we explore a variety of machine learning approaches to improve star-galaxy classification in ground-based photometric surveys. In Chapter 2, we present a meta-classification framework that combines existing star-galaxy classifiers, and demonstrate that our Bayesian combination technique improves the overall performance over any individual classification method. In Chapter 3, we show that a deep learning algorithm called convolutional neural networks is able to produce accurate and well-calibrated classifications by learning directly from the pixel values of photometric images. In Chapter 4, we study another deep learning technique called generative adversarial networks in a semi-supervised setting, and demonstrate that our semi-supervised method produces competitive classifications using only a small amount of labeled examples.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Physics

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2018

Author dc:creator

• Kim, Junhyung

Contributors dc:contributor

• Brunner, Robert J.
• Thaler, Jon J.
• Gollin, George D.
• Schwing, Alexander G.

• data analysis
• image processing
• photometric surveys
• star-galaxy classification
• cosmology
• deep learning
• convolutional neural networks
• generative adversarial networks

Statement dc:rights

• Copyright 2018 Junhyung Kim

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/100895

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/100895