Representation learning is a fundamental pillar of artificial intelligence, enabling models to extract and encode meaningful patterns from complex data into compact and informative representations. As a driving force behind the success of deep learning, effective representation learning empowers a wide array of applications, from computer vision and natural language processing to speech recognition and reinforcement learning, thereby advancing the capabilities of intelligent systems and their impact on society. We have gained significant progress by including big and curated data in training Deep Neural Networks (DNNs). However, the increasing labeling demand is expensive and time-consuming. As an alternative, semi-supervised or unsupervised learning approaches could generate high-quality representations. In the first topic of this dissertation, we focus on learning representations in a semi-supervised or self-supervised manner. Our research introduces a semi-supervised two-stage model that enables direct learning from noisy labels and facilitates the acquisition of high-quality representations for image classification tasks. Additionally, we suggest the utilization of a self-supervised model for addressing the audio-visual speaker diarization problem with an improved loss function. In the next topic, we explore the transferability of pre-trained DNNs to downstream tasks. We propose a relation transfer architecture for achieving domain adaptation in the referring expression grounding problem. We further investigate the robustness of learned representations and propose to prevent adversarial attacks on DNNs with attack-defendable neural network architectures. Finally, we calibrate the outputs of deep neural networks to improve the quality of uncertainty assessments. The dissertation also compares the proposed methods with other state-of-the-art approaches in the experiments. Overall, our research has the potential to enhance the efficiency, transferability, interpretability, and security of DNNs, contributing to the development of more powerful and trustworthy artificial intelligence systems.


If this is your thesis or dissertation, and want to learn how to access it or for more information about readership statistics, contact us at STARS@ucf.edu.

Graduation Date





Wang, Liqiang


Doctor of Philosophy (Ph.D.)


College of Engineering and Computer Science


Computer Science

Degree Program

Computer Science


CFE0009853; DP0028132





Release Date

November 2024

Length of Campus-only Access

1 year

Access Status

Doctoral Dissertation (Campus-only Access)

Restricted to the UCF community until November 2024; it will then be open access.