Continual Learning Self-Organising Maps with Controlled Plasticity

Abstract

Abstract Continual learning in neural networks is a crucial characteristic that mirrors the lifelong adaptability seen in human cognition. The Continual Learning Self-Organizing Map (CLSOM) modifies the traditional Self-Organizing Map (SOM) by replacing its time-dependent adaptation function with a time-invariant one. This alteration enables the CLSOM to support continual learning, similar to human cognitive processes. The model allows for setting the plasticity level on a continuum from absolute stability to absolute plasticity, providing a mechanism to control the stability-plasticity dilemma. Furthermore, drawing on human neuroscience insights, we introduce a variant of CLSOM featuring an “early plasticity surge” method to manage the challenges of random initialization and disrupted topology commonly found in standard SOMs. This approach mimics critical periods of enhanced plasticity in human neural development, aiming to improve the initial self-organization of the network to better adapt to dynamic environments.