Abstract

Most existing machine learning models assume that training and test data are independently and identically distributed (\iid) according to the same data distribution. Under this assumption, a model trained on labeled data is expected to generalize well to unseen test data. However, real-world applications often violate this assumption due to distributional shifts between the source and target domains (e.g., variations in lighting conditions in images or differences in writing styles in text). This research focuses on domain adaptation (DA)—a paradigm that aims to adapt models trained on one or more source domains to perform well on a distributionally different target domain, despite constraints on data availability and computing resources. Traditional DA methods assume access to labeled source domain data and unlabeled target domain data. However, in many real-world settings, these resources are often unavailable. For instance, privacy concerns restrict direct access to medical and financial data, while in autonomous driving, the vast diversity of environments—spanning different weather conditions, lighting scenarios, and geographic locations—makes it impractical to collect and label data for all possible object categories. In such cases, models must adapt to new conditions without additional labeling efforts, operating under resource-constrained settings where labeled source data or target supervision is limited or entirely absent. Under these constraints, robustness becomes a key challenge due to data corruption, unseen target categories, and model overconfidence. For instance, models trained on noisy source data risk error accumulation, failing to extract reliable domain-invariant features for generalization. The absence of prior knowledge about novel target classes makes it difficult to differentiate between known and unknown categories, leading to poor adaptation in open-set settings. And models trained on a different distribution tend to be overconfident in incorrect predictions, particularly for out-of-distribution samples, reducing reliability. To address these challenges naturally arising in many practical domain adaptation settings, this research aims to develop a robust domain adaptation framework that integrate noise-robust learning, open-set adaptation, and calibrated training to improve reliability under resource constrained conditions. It integrates the following key research components: (1) \textbf{Robust Domain adaptation with Noisy Labels}: Collection of clean labeled data is time-consuming and expensive. We study a robust domain adaptation framework under the sparsely labeled domains with corruptions in the context of few-shot learning and meta-learning. (2) \textbf{Multi-source domain adaptation with unlabeled data}: Extending domain adaptation to a multi-source setting, we incorporate an additional unlabeled source domain to mitigate reliance on labeled data. We introduce an unsupervised few-shot task and a noisy task filtration criterion to extract meaningful information from unlabeled and noisy samples. (3) \textbf{Source-free domain adaptation with pre-trained models}: When source data is inaccessible due to privacy or storage constraints, we enable adaptation using only pre-trained models. Our framework integrates knowledge from multiple pre-trained sources while mitigating error accumulation from noisy pseudo-labels. (4) \textbf{Open-set domain adaptation with a limited label space}: When the source domain has a restricted label set and the target domain includes unseen categories, we propose a domain-adaptive class-aware prompt to facilitate adaptation to both shared and novel categories, improving recognition in open-set settings. (5) \textbf{Calibrated fully test-time adaptation}: Due to the lack of supervision during adaptation, the source model’s predictions often become overconfident or under-confident. We design a calibration strategy to maintain reliable predictions during test-time adaptation. By systematically addressing these key challenges, this study provides a robust framework for domain adaptation in resource-constrained settings, enhancing model adaptability in scenarios where labeled data is scarce, noisy, or inaccessible. Our work contributes to the broader goal of making machine learning models more generalizable, efficient, and robust in real-world applications.

Library of Congress Subject Headings

Transfer learning (Machine learning); Computer vision--Data processing; Remote-sensing images--Classification

Publication Date

10-2025

Document Type

Dissertation

Student Type

Graduate

Degree Name

Computing and Information Sciences (Ph.D.)

Department, Program, or Center

Computing and Information Sciences Ph.D, Department of

College

Golisano College of Computing and Information Sciences

Advisor

Xumin Liu

Advisor/Committee Member

Qi Yu

Advisor/Committee Member

Daniel Krutz

Comments

This dissertation has been embargoed. The full-text will be available on or around 10/23/2026.

Campus

RIT – Main Campus

Plan Codes

COMPIS-PHD

Download

Available for download on Friday, October 23, 2026

Share

COinS