Imbalanced-Free Memory Selection Scheme Based Continual Learning by Using K-means Clustering

Abstract

Recently, continual learning is a new training strategy based on replaying memory and allowing beneficial effects to previous tasks. To preserve old information, current continual learning scheme accumulates observed examples into limited buffer or repeatedly trains generative model. This idea of learning scheme is effective to reduce catastrophic forgetting which is deterioration in overall performance when training sequentially. However, there still exists the problem of imbalanced data distribution in limited buffer and it is hard to apply on real-time system due to too long time to train both generative model and classification model. In this paper, we propose sample selection algorithm based on iterative k-means algorithm to improve the memory based continual learning. This approach selects examples to store into a buffer in a unsupervised manner using k-means cluster information. Our experiments on a variant of MNIST and CIFAR-100 datasets show the effects on classification accuracy performance when compared to the state-of-the-art.