Data-Efficient and Generalizable Machine Learning in Complex Environments
Access status:
Open Access
Type
ThesisThesis type
Doctor of PhilosophyAuthor/s
Xia, XiaoboAbstract
In an age marked by an unprecedented influx of data across diverse domains, the quest for effective machine learning (ML) solutions has increased significantly. However, data imperfections in complex environments present formidable obstacles, encompassing defective, redundant, and ...
See moreIn an age marked by an unprecedented influx of data across diverse domains, the quest for effective machine learning (ML) solutions has increased significantly. However, data imperfections in complex environments present formidable obstacles, encompassing defective, redundant, and scarce data. Specifically, defective data, characterized by annotation errors and incompleteness, obstruct the learning process, particularly in critical domains such as healthcare and finance. Redundant data overwhelm relevant insights, demanding efficient filtering techniques for optimal ML performance. Besides, scarce data that are prevalent in domains with limited examples, necessitate robust ML models capable of generalizing effectively. Addressing these challenges is pivotal for unlocking the full potential of ML technologies. This thesis offers innovative solutions across three key areas: learning with defective data, redundant data, and scarce data. Particularly, for defective data, it explores learning with mislabelled and incomplete data, which proposes novel methods for handling each scenario. In the realm of redundant data, the thesis introduces a moderate coreset selection technique to enhance ML efficiency across diverse practical tasks, and a refined coreset selection strategy to reduce the size of the constructed coreset while maintaining satisfactory model performance. Additionally, it addresses the challenge of scarce data by proposing advanced strategies for kernel mean estimation and augmenting datasets by marginalized corruption distributions to improve sample efficiency and model generalization. This thesis provides comprehensive insights and solutions for learning with imperfect data. By addressing these obstacles, it promotes the development of data-efficient and generalizable ML, and lays the groundwork for transformative breakthroughs in fields such as healthcare, finance, and climate science, propelling innovation and progress fuelled by the power of ML.
See less
See moreIn an age marked by an unprecedented influx of data across diverse domains, the quest for effective machine learning (ML) solutions has increased significantly. However, data imperfections in complex environments present formidable obstacles, encompassing defective, redundant, and scarce data. Specifically, defective data, characterized by annotation errors and incompleteness, obstruct the learning process, particularly in critical domains such as healthcare and finance. Redundant data overwhelm relevant insights, demanding efficient filtering techniques for optimal ML performance. Besides, scarce data that are prevalent in domains with limited examples, necessitate robust ML models capable of generalizing effectively. Addressing these challenges is pivotal for unlocking the full potential of ML technologies. This thesis offers innovative solutions across three key areas: learning with defective data, redundant data, and scarce data. Particularly, for defective data, it explores learning with mislabelled and incomplete data, which proposes novel methods for handling each scenario. In the realm of redundant data, the thesis introduces a moderate coreset selection technique to enhance ML efficiency across diverse practical tasks, and a refined coreset selection strategy to reduce the size of the constructed coreset while maintaining satisfactory model performance. Additionally, it addresses the challenge of scarce data by proposing advanced strategies for kernel mean estimation and augmenting datasets by marginalized corruption distributions to improve sample efficiency and model generalization. This thesis provides comprehensive insights and solutions for learning with imperfect data. By addressing these obstacles, it promotes the development of data-efficient and generalizable ML, and lays the groundwork for transformative breakthroughs in fields such as healthcare, finance, and climate science, propelling innovation and progress fuelled by the power of ML.
See less
Date
2024Rights statement
The author retains copyright of this thesis. It may only be used for the purposes of research and study. It must not be used for any other purposes and may not be transmitted or shared with others without prior permission.Faculty/School
Faculty of Engineering, School of Computer ScienceAwarding institution
The University of SydneyShare