Files and links (1)
Published, Version of Record (VoR)Open Access via Drexel Libraries Read and Publish Program 2026CC BY V4.0, Open
Journal article
Robust incremental learning of visual concepts without catastrophic forgetting
Cognitive systems research, v 96, 101447
Apr 2026
Featured in Collection : Drexel's Newest Publications
Abstract
Continual learning poses a significant challenge in machine learning, as models often struggle to retain previously learned knowledge when exposed to new data, leading to catastrophic forgetting. In this work, we introduce Cobweb/4V, a novel visual classification method. This approach builds on Cobweb, a human-like learning system that is inspired by the way humans incrementally learn new concepts over time. In this research, we conduct a comprehensive evaluation showcasing Cobweb/4V’s proficiency in learning visual concepts, requiring less data to achieve effective learning outcomes compared to neural network approaches, maintaining stable performance over time, achieving competitive asymptotic behavior, and avoiding catastrophic forgetting. These characteristics align with human learning capabilities, positioning Cobweb/4V as a promising approach for sequential learning and motivating future exploration into its potential to guide the development of neural networks and other machine learning approaches that handle continual learning.
Metrics
6 Record Views
Details
- Title
- Robust incremental learning of visual concepts without catastrophic forgetting
- Creators
- Nicki Barari (Corresponding Author) - Drexel UniversityXin Lin - Northwestern UniversityChristopher J. MacLellan - Georgia Institute of Technology
- Publication Details
- Cognitive systems research, v 96, 101447
- Publisher
- Elsevier
- Number of pages
- 8
- Grant note
- NSF's AI ALOE institute and Awards: 2112532 U.S. Army Research Laboratory (ARL), STRONG Program: W911NF2120101, W911NF2320203, W911NF2120126
This research was partially funded by Award 2112532 from NSF's AI-ALOE institute and Awards W911NF2120101, W911NF2320203, and W911NF2120126 from ARL's STRONG program. The views, opinions, and findings expressed are the authors' and should not be taken as representing the official views or policies of these funding agencies. Additionally, we extend our sincere appreciation to individuals who generously shared their invaluable insights and feedback on this work, notably Pat Langley (GTRI), Jesse Roberts (Tennessee Tech), Kyle Moore (Vanderbilt), and Doug Fisher (Vanderbilt).
- Resource Type
- Journal article
- Language
- English
- Academic Unit
- Computer Science
- Web of Science ID
- WOS:001681556100002
- Other Identifier
- 991022155069604721