Abstract
The rapid evolution of AI heightens the need for learning systems that are efficient, robust, and explainable. This dissertation advances these three pillars through innovations in classification, generative modeling, domain adaptation under data-scarce conditions, and multimodal retrieval. Collectively, the methods reduce dependence on large, labeled datasets, improve adaptability under distribution shifts, enable deployment on resource-constrained platforms, and enhance interpretability. For classification, the Iterative Maximum Likelihood Classifier (IMLC) recasts regularized maximum likelihood training as a fixed-point contraction with convergence guarantees, enabling faster and more stable optimization. Results on synthetic data and MNIST validate its efficiency. For generative models, we introduce Parametric Mish (PMish) activation, a modified MMD-GAN repulsive loss, and Adaptive Rank Decomposition (ARD) for compression. Together, these stabilize training, improve convergence, and reduce model size while maintaining fidelity. Benchmarks (CIFAR-10/100, STL-10, CelebA) show strong image quality with lower resources. Moving from model efficiency to data robustness, two data augmentation methods, namely Shuffle PatchMix (SPM) and Dual-Region Augmentation (DRA), are proposed to improve model robustness in data-scarce scenarios. Using these augmentations, along with novel loss functions and confidence-margin reweighting for noisy pseudo-labels, yields consistent gains under distribution shift. For multimodal retrieval, X-CoT combines coarse video-text matching with large language model chain-of-thought reasoning to raise retrieval accuracy and produce human-interpretable rationales. Together, these contributions demonstrate that performance, adaptability, and transparency can be achieved, laying a foundation for next-generation machine learning systems that are efficient, robust, and explainable.
Publication Date
12-2025
Document Type
Dissertation
Student Type
Graduate
Degree Name
Electrical and Computer Engineering (Ph.D)
College
Kate Gleason College of Engineering
Advisor
Majid Rabbani
Advisor/Committee Member
Sohail Dianat
Advisor/Committee Member
Jamison Heard
Recommended Citation
Pulakurthi, Prasanna Reddy, "Advancements in ML via Efficient Generative Modeling, Robust Domain Adaptation, and Explainable Multimodal Retrieval" (2025). Thesis. Rochester Institute of Technology. Accessed from
https://repository.rit.edu/theses/12374
Campus
RIT – Main Campus
