Over the last decade, we have witnessed the renaissance of deep neural networks (DNNs) and their successful applications in computer vision. There is still a long way to build intelligent and reliable machine vision systems, but DNNs provide a promising direction. The goal of this thesis is to present a few small steps along this road. We mainly focus on two questions: How to design label-efficient learning algorithms for computer vision tasks? How to improve the robustness of DNN based visual models? Concerning label-efficiency, we investigate a reinforced sequential model for video summarization, a background hallucination strategy for high-resolution image generation, and a selective module integrated into self-supervised self-training for improving object detection with noisy Web images. Besides, we study how to rank many pre-trained deep neural checkpoints for the transfer learning to a downstream task. Considering robustness, we propose a powerful blackbox adversarial attack to facilitate the research toward robust DNNs, and we also explore a new threat model that the adversaries can distill the knowledge from a blackbox teacher model to harvest a student model for imitating the characteristics of the teacher. In each chapter, we introduce the problem and present our solutions using machine learning and deep neural architectures, followed by comparisons with existing baselines and discussions on future research.
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.
Doctor of Philosophy (Ph.D.)
College of Engineering and Computer Science
Length of Campus-only Access
Doctoral Dissertation (Open Access)
Li, Yandong, "Learning Accurate and Robust Deep Visual Models" (2021). Electronic Theses and Dissertations, 2020-. 521.