In recent years, progress in computing and networking has made it possible to collect large volumes of data for various different applications in data mining and data analytics using machine learning methods. Data may come from different sources and in different shapes and forms depending on their inherent nature and the acquisition process. In this dissertation, we focus specifically on sequential data, which have been exponentially growing in recent years on platforms such as YouTube, social media, news agency sites, and other platforms. An important characteristic of sequential data is the inherent causal structure with latent patterns that can be discovered and learned from samples of the dataset. With this in mind, we target problems in two different domains of Computer Vision and Natural Language Processing that deal with sequential data and share the common characteristics of such data. The first one is action recognition based on video data, which is a fundamental problem in computer vision. This problem aims to find generalized patterns from videos to recognize or predict human actions. A video contains two important sets of information, i.e. appearance and motion. These information are complementary, and therefore an accurate recognition or prediction of activities or actions in video data depend significantly on our ability to extract them both. However, effective extraction of these information is a non-trivial task due to several challenges, such as viewpoint changes, camera motions, and scale variations, to name a few. It is thus crucial to design effective and generalized representations of video data that learn these variations and/or are invariant to such variations. We propose different models that learn and extract spatio-temporal correlations from video frames by using deep networks that overcome these challenges. The second problem that we study in this dissertation in the context of sequential data analysis is text summarization in multi-document processing. Sentences consist of sequence of words that imply context. The summarization task requires learning and understanding the contextual information from each sentence in order to determine which subset of sentences forms the best representative of a given article. With the progress made by deep learning, better representations of words have been achieved, leading in turn to better contextual representations of sentences. We propose summarization methods that combine mathematical optimization, Determinantal Point Processes (DPPs), and deep learning models that outperform the state of the art in multi-document text summarization.
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)
Cho, Sangwoo, "Contextual Understanding of Sequential Data Across Multiple Modalities" (2021). Electronic Theses and Dissertations, 2020-. 483.