University of North Texas
Latent-Space Engineering and Modular Transformer Architectures for Efficient Multimodal Intelligence
Abstract
dc:descriptionThis thesis explores advanced methodologies in machine learning and latent-space engineering, emphasizing the critical role of architecture modularity in optimizing resource-constrained artificial intelligence. It presents a systematic investigation of a lightweight, transformer-inspired engine designed to achieve high-throughput inference across heterogeneous signal domains without the prohibitive computational overhead typical of modern Vision Transformers (ViTs). Chapter 3 establishes the foundational Binary Classification Vision Transformer (BC-ViT), integrating convolutional feature extraction with transformer-based contextual reasoning to excel in low-visibility underwater object classification. Chapter 4 extends this baseline int the Waste Management Classification Vision Transformer (WMC-ViT), employing parallel multi-head attention to achieve highly efficient multi-class environmental reasoning, demonstrating a 99.74\% reduction in parameters compared to state-of-the-art architectures. Chapter 5 introduces the Edge IR Vision Transformer (EIR-ViT), adapting the modular framework to process thermal intensity gradients for infrared target segmentation in tactical environments. Chapter 6 explores temporal sequence modeling, integrating Long Short-Term Memory (LSTM) layers to create a hybrid CNN-LSTM-transformer architecture capable of nonlinear energy consumption forecasting. Chapter 7 and 8 demonstrate the architecture's robust adaptability within clinical medical imaging. The CXR8-ViT handles multi-label thoracic pathology classification from chest radiographs, while the BT-ViT provides adaptive solutions for high-complexity CT and MRI-based brain tumor classification with varying dimensionality sizes. Chapter 9 extends these advancements to remote sensing through the Multi-Spectral Efficient Land Use Classification (MELC and MELC-PC) models, utilizing low-rank convolutional filtering for planetary-scale satellite imagery. Chapter 10 introduces the Sawzall framework, a unified spatial-spectral fusion architecture that successfully disparate optical, radar, and multispectral data streams into a single representational manifold. This work underscores the transformative impact of coupling lightweight neural architectures with domain-agnostic modularity, paving the way for efficient, multimodal artificial intelligence deployable directly at the edge.
Degree
thesis:*- Grantor dc:publisher
- University of North Texas
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Kurz, Aidan Gabriel 2000-
- Contributors dc:contributor
-
- Park, Hyusim
- Bailey, Colleen P.
- Namuduri, Kamesh
Subjects
dc:subject × 3Rights
dc:rights- Statement dc:rights
-
- Public
- Kurz, Aidan Gabriel 2000-
- Copyright
- Copyright is held by the author, unless otherwise noted. All rights Reserved.
- Language dc:language
- English
Identifiers
dc:identifier.*- Identifier
-
local-cont-no: submission_5042
ark: ark:/67531/metadc2652535