Theses and Dissertations
Advisor
Mohammadi-Aragh, Jean
Committee Member
Ball, John
Committee Member
Kennedy, Ken
Committee Member
Ma, Junfeng
Date of Degree
5-15-2026
Original embargo terms
Embargo 1 year
Document Type
Dissertation - Open Access
Major
Computational Engineering
Degree Name
Doctor of Philosophy (Ph.D.)
College
James Worth Bagley College of Engineering
Department
Computational Engineering Program
Abstract
Modern deep learning performance is often attributed to model architecture or scale, yet the role of dataset structure remains poorly formalized and difficult to consolidate. This dissertation addresses this gap through three studies that use controlled synthetic data and statistical analysis to characterize how data composition influences convolutional neural networks (CNNs) in terms of learning, generalization, and robustness. The first study introduces a programmatic synthetic data framework that systematically varies object-level and non-object attributes while holding all other features constant. Using multiple CNN architectures trained on these controlled datasets, the study quantifies how individual feature groups contribute to performance and demonstrates that CNNs implicitly prioritize statistically frequent and structurally consistent patterns, aligning with Apriori principles from frequent pattern mining. Building on these findings, the second study develops a dataset evaluation and subset optimization methodology using PCA, LDA, and Shapley inspired contribution analysis. This study shows that carefully selected subsets can preserve or improve model performance while substantially reducing dataset size, and that dataset level structure provides stronger explanatory power than individual sample difficulty alone. The final study extends these insights through a dynamic training system, DITTO, a structure aware weighting framework that adapts optimization pressure based on observed data imbalance and feature impact. Across multiple MNIST benchmark datasets, DITTO improves performance metrics and training stability through anchored loss based adjustments.
Recommended Citation
O'Sullivan, Eric William, "Anchoring object features through controlled synthetic data" (2026). Theses and Dissertations. 6957.
https://scholarsjunction.msstate.edu/td/6957
