Abstract
Continual learning poses a fundamental challenge for neural systems, which typically suffer from catastrophic forgetting when exposed to sequential tasks. Self-Organizing Maps (SOMs), despite their inherent interpretability and efficiency, also exhibit this vulnerability. In this paper, we introduce Saturation Self-Organizing Maps (SatSOM)—an extension designed to enhance knowledge retention in continual learning scenarios. Sat-SOM incorporates a novel saturation mechanism that progressively reduces the learning rate and neighborhood radius of neurons as they accumulate information. This dynamic effectively stabilizes well-trained neurons, redirecting new learning to underutilized regions of the map. To further accommodate tasks of unknown complexity, we introduce a dynamic variant capable of adaptive grid expansion. We evaluate SatSOM on sequential versions of the FashionMNIST and KMNIST datasets, showing that it significantly outperforms existing SOM-based methods and approaches the retention capabilities of a k-nearest neighbors (kNN) baseline. Ablation studies confirm the critical role of the saturation mechanism. SatSOM offers a lightweight and interpretable solution for sequential learning and provides a foundation for implementing adaptive plasticity in complex architectures.