Electronic Theses and Dissertations

Sahil Nokhwal

Date

2025

Document Type

Dissertation

Degree Name

Doctor of Philosophy

Department

Computer Science

Committee Chair

Sajjan G. Shiva

Committee Member

Ching-Chi Yang

Committee Member

Deepak Venugopal

Committee Member

Nirman Kumar

Abstract

Neural network (NN) models have made remarkable strides in outperforming humans in a wide array of discrete tasks. These models excel in specific, well-defined domains where they can leverage vast amounts of data to make highly accurate predictions. However, despite impressive achievements, the scope of NN remains limited when compared to the vast, multifaceted cognitive abilities of humans. While machine learning models excel at specialized tasks, humans possess an extraordinary capacity to acquire knowledge and perform a nearly infinite variety of tasks across diverse domains. The ability to learn over time is a core attribute of human cognition. This lifelong process allows humans to integrate new information while retaining prior knowledge. In contrast, machine learning models struggle, especially in dynamic environments. Continual learning (CL) addresses this challenge, enabling models to learn from a stream of data, whether online or offline, without forgetting past knowledge. A critical challenge in CL is the phenomenon known as "catastrophic forgetting" (CF). CF occurs when an NN, after learning new information, forgets previously acquired knowledge, often leading to a degradation in performance on earlier tasks. This problem has been extensively studied and remains one of the most significant hurdles in deep neural networks (DNNs). In this dissertation, the focus has been on presenting novel solutions to mitigate CF in the continual and incremental learning (CL and IL) domain. The solutions proposed in this dissertation address several crucial aspects of CL. By exploring different strategies and techniques, the research advances the field by addressing key challenges such as balancing the retention of old knowledge while incorporating new data. These contributions reflect significant progress in the development of CL techniques and underscore the obstacles that still need to be overcome for the effective implementation of CL in real-world applications. Through these efforts, this research aims to enhance the practical viability of CL and make it more applicable to complex, dynamic environments.

Comments

Data is provided by the student.

Library Comment

Dissertation or thesis originally submitted to ProQuest.

Notes

Open access

Download

Share

COinS