Dynamic learning — An approach to forgetting in ART2 neural networks

Abstract

In machine learning “forgetting” little used or redundant information can be seen as a sensible strategy directed at the overall management of specific and limited computational resources. This paper describes new learning rules for the ART2 neural network model of category learning that facilitates forgetting without additional node features or subsystems and which preserves the main characteristics of the classic ART2 model. We consider that this approach is straightforward and is arguably biological plausible. The new learning rules drop the specification within the classic ART2 model that learning should only occur at the winning node. Classic ART2 learning rules are presented as a particular case of these new rules. The model increases system adaptability to continually changing or complex input domains. This allows the system to maintain information in a manner which is consistent with its use and allows system resources to be dynamically allocated in a way that is consistent with observations made of biological learning.