Overcoming Catastrophic Forgetting - Geometric Techniques in Incremental Machine Learning

Aug 4, 2026

Upload by Sage

Neural network models achieve strong performance on discrete tasks but struggle to match humans’ lifelong, multifaceted learning. Continual learning addresses this gap by enabling models to learn from data streams while retaining prior knowledge. A central difficulty is catastrophic forgetting, where learning new information degrades performance on earlier tasks. This dissertation proposes novel solutions for continual and incremental learning, focusing on balancing old knowledge retention with integration of new data.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages187

Reading Time7 h 51 min

AccessFree to download

Acknowledgments

Abstract

List of Publications

University of Memphis University of Memphis Digital Commons Electronic Theses and Dissertations 5-9-2025 Overcoming Catastrophic Forgetting: Geometric Techniques in Incremental Machine Learning Sahil Nokhwal Follow this and additional works at: [https://digitalcommons.memphis.edu/etd](https://digitalcommons.memphis.edu/etd) Recommended Citation Nokhwal, Sahil, "Overcoming Catastrophic Forgetting: Geometric Techniques in Incremental Machine Learning" (2025) . Electronic Theses and Dissertations. 3802. [https://digitalcommons.memphis.edu/etd/3802](https://digitalcommons.memphis.edu/etd/3802) This Dissertation is brought to you for free and open access by University of Memphis Digital Commons. It has been accepted for inclusion in Electronic Theses and Dissertations by an authorized administrator of University of Memphis Digital Commons. For more information, please contact [khggerty@memphis.edu](khggerty@memphis.edu). OVERCOMING CATASTROPHIC FORGETTING: GEOMETRIC TECHNIQUES IN INCREMENTAL MACHINE LEARNING by Sahil Nokhwal A Dissertation Submitted in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy Major: Computer Science Committee Members: Sajjan G. Shiva, Ph.D. (Committee Chair) Nirman Kumar, Ph. D. Deepak Venugopal, Ph.D. Ching-Chi Yang, Ph.D. The University of Memphis May 2025 © Copyright 2025 Sahil Nokhwal Partial rights reserved ii ACKNOWLEDGMENTS This dissertation is the culmination of years of hard work, and it would not have been possible without the unwavering support of numerous individuals andinstitutions. I have been privileged to work under the guidance of Dr. Sajjan G. Shiva and Dr. Nirman Kumar, whose expertise, mentorship, and constant encouragement have been essential to the completion of this research. I am deeply grateful for their insightful feedback and the direction they provided throughout my journey. I would also like to extend my gratitude to the faculty and staff at the University of Memphis, especially my dissertation committee members, whose valuable insights have enriched my work. The stimulating discussions and collaborative atmosphere provided by my colleagues, both at the university and beyond, have played a crucial role in shaping this dissertation. I am also very thankful for the spiritual and religious institutions which have had a profound impact and provided strength throughout this PhD journey. Finally, to my friends and family—thank you for your unconditional love, support, and belief in me. Your encouragement has been my pillar, and I could not have made it this far without you. To everyone who has supported me in any way throughout this journey, my heartfelt thanks. 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 anearly 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

FAQs

What challenge does continual learning address?

Continual learning enables models to learn from a stream of data without forgetting previously learned knowledge, especially in dynamic environments.

What is catastrophic forgetting in incremental/continual learning?

Catastrophic forgetting occurs when, after learning new information, a neural network forgets previously acquired knowledge, degrading performance on earlier tasks.

How does this dissertation aim to mitigate catastrophic forgetting?

It presents novel solutions for continual and incremental learning by exploring strategies and techniques that balance retaining old knowledge while incorporating new data.

Aug 4, 2026

Upload by Sage