Continual Recognition with Adaptive Memory Update

Class incremental continual learning aims to improve the ability of modern classification models continually recognize new classes without forgetting previous ones. Prior art in field has largely considered using a replay buffer. In this article, we start from an observation that existing replay-based method would fail when stored exemplars are not hard enough get good decision boundary between previously learned class and class. To prevent situation, propose perspective remedy after for first time. proposed method, set is preserved as working memory, which helps classes. When memory insufficient distinguish classes, more discriminating samples be swapped long-term built up during early training process, adaptive way. Our recognition model with update capable overcoming problem catastrophic various coming sequence, especially similar but different Extensive experiments on real-world datasets demonstrate superior state-of-the-art algorithms. Moreover, our can used general plugin any algorithm further their performance.