Abstract

Modern AI systems can achieve remarkable performance on tasks such as image recognition, object detection, and language understanding, but they often struggle with a basic ability that humans rely on every day: learning continuously. When deep neural networks (DNNs) are updated with new information, they can unexpectedly forget what they previously learned, making them difficult to deploy in changing real-world environments. This dissertation explores how to build AI systems that can learn more like humans: adapting to new data, preserving past knowledge, and using prior experience to learn future tasks more efficiently. Rather than focusing only on preventing forgetting, this work also addresses two critical requirements for practical continual learning (CL): computational efficiency and forward transfer. The dissertation introduces several methods that improve how DNNs learn from evolving data streams, including biologically inspired wake/sleep training, efficient rehearsal strategies, techniques for reducing stability gap, and approaches for improving out-of-distribution (OOD) generalization (i.e., forward transfer) and detection. Together, these contributions move CL closer to real-world deployment by making AI systems more efficient, adaptable, and reliable in open-world environments. The broader goal of this work is to help develop machine learning systems that can keep improving over time without the costly need to repeatedly retrain from scratch.

Publication Date

6-8-2026

Document Type

Dissertation

Student Type

Graduate

Degree Name

Imaging Science (Ph.D.)

Department, Program, or Center

Chester F. Carlson Center for Imaging Science

College

College of Science

Advisor

Christopher Kanan,

Advisor/Committee Member

Ferat Sahin,

Advisor/Committee Member

Nathan Cahill

Campus

RIT – Main Campus

Download

Share

COinS