Date of Award
4-2021
Access Type
Dissertation - Open Access
Degree Name
Doctor of Philosophy in Electrical Engineering & Computer Science
Department
College of Engineering
Committee Chair
Houbing Song, Ph.D.
First Committee Member
Radu F. Babiceanu, Ph.D.
Second Committee Member
Richard S. Stansbury, Ph.D.
Third Committee Member
Tianyu (Thomas) Yang, Ph.D.
Fourth Committee Member
Xingquan (Hill) Zhu, Ph.D.
Abstract
Machine learning (ML) has attracted a significant amount of attention from the artifi- cial intelligence community. ML has shown state-of-art performance in various fields, such as signal processing, healthcare system, and natural language processing (NLP). However, most conventional ML algorithms suffer from three significant difficulties: 1) insufficient high-quality training data, 2) costly training process, and 3) domain dis- crepancy. Therefore, it is important to develop solutions for these problems, so the future of ML will be more sustainable. Recently, a new concept, data-efficient ma- chine learning (DEML), has been proposed to deal with the current bottlenecks of ML. Moreover, transfer learning (TL) has been considered as an effective solution to address the three shortcomings of conventional ML. Furthermore, TL is one of the most active areas in the DEML. Over the past ten years, significant progress has been made in TL.
In this dissertation, I propose to address the three problems by developing a software- oriented framework and TL algorithms. Firstly, I introduce a DEML framework and a evaluation system. Moreover, I present two novel TL algorithms and applications on real-world problems. Furthermore, I will first present the first well-defined DEML framework and introduce how it can address the challenges in ML. After that, I will give an updated overview of the state-of-the-art and open challenges in the TL. I will then introduce two novel algorithms for two of the most challenging TL topics: distant domain TL and cross-modality TL (image-text). A detailed algorithm introduction and preliminary results on real-world applications (Covid-19 diagnosis and image clas- sification) will be presented. Then, I will discuss the current trends in TL algorithms and real-world applications. Lastly, I will present the conclusion and future research directions.
Scholarly Commons Citation
Niu, Shuteng, "Data-Efficient Machine Learning with Focus on Transfer Learning" (2021). Doctoral Dissertations and Master's Theses. 574.
https://commons.erau.edu/edt/574