Continual Learning with Self-Organizing Maps: A Novel Group-Based Unsupervised Sequential Training Approach

Continual learning is essential for intelligent systems to thrive. While deep learning models based on Convolutional Neural Networks are robust, they can be complex and resource-intensive, requiring a long and computationally intensive training phase due to Back-propagation (BP). On the other hand, unsupervised methods are favored for their lightweight, speed, and ease of management, making them ideal for ongoing learning. Self-organizing maps (SOMs) have gained attention in computer vision over the past decade due to their excellent feature extraction abilities and transparency. However, SOMs encounter challenges in learning, such as having a fixed map size that limits new feature accommodation and the risk of interference when retraining a pre-trained map with the latest data. Although some studies have explored using SOMs to tackle these issues, obstacles like catastrophic forgetting and lifelong learning hurdles remain prevalent. To combat forgetting issues, we propose an approach called learning by grouping. In this method, each group generates a trained map stored in memory before the model is reset and trained on the group to create another map. This sequential process facilitates learning while preserving previously acquired knowledge without disruptions. The suggested approach has been tried out on MNIST digit, Fashion MNIST, and CIFAR 10 datasets. The proposed method demonstrates final accuracy improvement of 60-70% across all three datasets when compared to traditional SOM, indicating the effectiveness of grouping in overcoming the limitations of SOM-based continual learning methods.