Incremental Concept Formation over Visual Images Without Catastrophic Forgetting

TL;DR

This work presents Cobweb/4V, a human-inspired, tensor-based visual classification method that incrementally forms concepts to mitigate catastrophic forgetting in continual learning. By switching to an information-theoretic category utility, employing a multi-concept prediction strategy, and representing images as pixel tensors, Cobweb/4V achieves data-efficient learning and robust retention on MNIST, outperforming neural baselines under forgetting scenarios. The experiments demonstrate rapid initial learning, competitive final accuracy, and significantly reduced forgetting compared with replay-based neural networks, highlighting the potential of modular, instance-based approaches for vision tasks. The study also outlines avenues to extend Cobweb with representation learning techniques (e.g., convolutional or attention mechanisms) to further boost performance while preserving memory stability.

Abstract

Deep neural networks have excelled in machine learning, particularly in vision tasks, however, they often suffer from catastrophic forgetting when learning new tasks sequentially. In this work, we introduce Cobweb4V, an alternative to traditional neural network approaches. Cobweb4V is a novel visual classification method that 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 Cobweb4Vs proficiency in learning visual concepts, requiring less data to achieve effective learning outcomes compared to traditional methods, maintaining stable performance over time, and achieving commendable asymptotic behavior, without catastrophic forgetting effects. These characteristics align with learning strategies in human cognition, positioning Cobweb4V as a promising alternative to neural network approaches.

Figures (6)

• Figure 1: A new instance is incorporated into an empty Cobweb tree. Cobweb adds it to the root and updates the root's count table to reflect the instance's attribute-values.
• Figure 2: Cobweb's Learning Process
• Figure 3: The single image on the left shows the tensor-based representation for instances. The concepts store the means and standard deviations for the pixel attributes in internal tensor representations as well.
• Figure 4: The maximum best nodes considered in predictions combination vs. the averaged test accuracy with 95% confidence intervals (evaluated on the entire MNIST test set) of Cobweb/4V after training on the entire MNIST training set.
• Figure 5: Average test accuracy with 95% confidence intervals on the MNIST test set as the number of training examples used increased in Experiment 1. We only present the learning curve of each approach after training with 1,000 examples.