Unsupervised Continual Learning Achieved with Growing Self-Organizing Maps

Continual learning systems face the challenge of learning new tasks without forgetting previously acquired knowledge, often requiring the storage of past data exemplars. This paper proposes a novel unsupervised framework that overcomes this limitation by employing growing self-organizing maps (GSOMs). These GSOMs are enhanced with distributional statistics and, optionally, encoder-decoder models. The core innovation lies in its ability to perform "exemplar-free replay." Instead of storing raw data, each GSOM unit maintains estimates of its own mean, variance, and covariance. These statistics are then used to generate synthetic samples for replay, which can be decoded back into the input space for subsequent training. This fully unsupervised approach, which doesn't require task boundaries or class labels, demonstrates performance competitive with state-of-the-art supervised memory-based methods across various benchmarks, particularly in challenging single-class incremental scenarios.