In the last half-decade, a new renaissance of machine learning originates from the applications of convolutional neural networks to visual recognition tasks. It is believed that a combination of big curated data and novel deep learning techniques can lead to unprecedented results. However, the increasingly large training data is still a drop in the ocean compared with scenarios in the wild. In this literature, we focus on learning transferable representation in the neural networks to ensure the models stay robust, even given different data distributions. We present three exemplar topics in three chapters, respectively: zero-shot learning, domain adaptation, and generalizable adversarial attack. By zero-shot learning, we enable models to predict labels not seen in the training phase. By domain adaptation, we improve a model's performance on the target domain by mitigating its discrepancy from a labeled source model, without any target annotation. Finally, the generalization adversarial attack focuses on learning an adversarial camouflage that ideally would work in every possible scenario. Despite sharing the same transfer learning philosophy, each of the proposed topics poses a unique challenge requiring a unique solution. In each chapter, we introduce the problem as well as present our solution to the problem. We also discuss some other researchers' approaches and compare our solution to theirs in the experiments.
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)
Zhang, Yang, "Learning Transferable Representations for Visual Recognition" (2020). Electronic Theses and Dissertations, 2020-. 162.
Restricted to the UCF community until May 2020; it will then be open access.