Graduation Semester and Year
Fall 2026
Language
English
Document Type
Dissertation
Degree Name
Doctor of Philosophy in Computer Science
Department
Computer Science and Engineering
First Advisor
Dr. Jacob Luber
Second Advisor
Dr. Manfred Huber
Third Advisor
Dr. William Beksi
Fourth Advisor
Dr. Junzhou Huang
Abstract
High-dimensional imaging data has become central to modern science, yet its scale and complexity make it difficult to model, compress, and interpret. This dissertation develops and applies machine learning methods that extract compact, informative, and predictive representations from such data, and demonstrates their reach across two very different domains: the reconstruction of natural images and the analysis of transcriptional dynamics during \textit{Drosophila} embryogenesis. Across three studies, a shared methodological thread---learning efficient representations of complex image data---connects advances in computer vision to the mechanistic study of gene regulation.
The first study addresses a core limitation of vector-quantized image models: poor codebook utilization, in which only a fraction of code vectors are used, capping reconstruction fidelity. It introduces a Dual Codebook mechanism that partitions the latent representation into complementary global and local components---the former refined by a lightweight transformer that updates all code vectors concurrently, the latter maintained through deterministic selection. Trained entirely from scratch without pretrained priors, the method achieves state-of-the-art reconstruction quality while using a codebook half the size of competing approaches, establishing that the utilization of a discrete latent space, not merely its size, governs reconstruction fidelity.
The second study extends representation learning from natural images to biological microscopy, developing a machine learning framework that predicts the spatiotemporal distribution of active nuclei along the dorsal--ventral axis of the embryo directly from super-resolution live-imaging data. The pipeline accurately recovers wild-type expression dynamics and proves sensitive to subtle genetic perturbation, distinguishing wild-type \textit{sogD} from the \textit{sogD}\,$\Delta$Su(H) mutant and detecting the broader domain of active nuclei that emerges when Su(H) binding is lost.
The third study applies quantitative single-nucleus live imaging to reveal that Su(H) plays a stage-dependent role in enhancer regulation that runs counter to its familiar function as a repressor. In the lead-up to gastrulation, the loss of Su(H) binding produces fewer transcripts, delayed initiation, and more frequent but shorter-lived transcriptional bursts, implying that Su(H) acts here to bolster robust transcription---a pro-expression role that is attenuated in earlier cycles and likely tied to the onset of Notch signaling.
Together, these studies show that machine learning can serve as a common foundation across otherwise unrelated imaging problems, advancing both high-fidelity image modeling and the quantitative understanding of how gene expression is regulated in space and time.
Keywords
Representation Learning, Image generation, Image Reconstruction, VQGAN
Disciplines
Other Electrical and Computer Engineering
License
This work is licensed under a Creative Commons Attribution 4.0 International License.
Recommended Citation
Boodaghimalidarreh, Parisa, "REPRESENTATION LEARNING: FROM COMPUTER VISION TO APPLICATIONS IN COMPUTATIONAL BIOLOGY" (2026). Computer Science and Engineering Dissertations. 25.
https://mavmatrix.uta.edu/cse_dissertations2/25