Semester of Graduation
Summer 2026
Degree Type
Dissertation
Degree Name
Doctor of Philosophy in Data Science and Analytics
Department
School of Data Science and Analytics - College of Computing and Software Engineering
Committee Chair/First Advisor
Christopher Cornelison
Second Advisor
Ramazan Aygun
Third Advisor
Paola Spoletini
Fourth Advisor
Michail Alexiou
Abstract
Deep learning has achieved broad success, yet many models remain limited by how effectively they preserve and use contextual information. Important relationships may be lost during regularization, transformed through fixed nonlinearities, distributed across multiple spatial scales, or remain only partially observable in long-horizon prediction. This dissertation investigates contextual representation learning as a framework for improving how such information is preserved, transformed, reconstructed, and integrated.
Four complementary studies address these challenges. Leaky Dropout introduces graded stochastic attenuation that preserves weak feature information while maintaining regularization. Morphos develops adaptive activation functions whose nonlinear behavior is learned from data rather than fixed in advance. A teacher-guided simulation framework reconstructs historical context for long-horizon sequence prediction and downstream optimization. Finally, the Hybrid Vision Transformer and Zoom-Out, Zoom-In framework integrate local and global visual information through hierarchical multiscale learning and adaptive supervision.
Across classification, sequential prediction, optimization, and segmentation tasks, the proposed methods show that improving contextual representations can enhance learning without relying solely on greater model capacity. The findings indicate that preserving information during optimization, adapting nonlinear transformations, reconstructing historical state, and integrating representations across scales are complementary strategies for context-sensitive learning. Collectively, this work suggests that future advances in machine learning may depend as much on constructing richer, more adaptive internal representations as on increasing architectural complexity.
Included in
Artificial Intelligence and Robotics Commons, Data Science Commons, Longitudinal Data Analysis and Time Series Commons, Theory and Algorithms Commons
Comments
This dissertation was supported in part by the Georgia Research Alliance (GRA) Phase IA Commercialization Grant (2025), "CellSlice.ai: AI-Powered Microbial Identification from Microscopy Images" ($25,000).