Capacity Control with Class Difficulty in Continual Learning of Image Classification Using Deep Learning

Abstract

In this study, we focus on incremental class continual learning for image classification, and improve the replay method that retains a part of the data given in the past and uses it to train a newly increased class. In the existing studies of replay methods, the number of data to be retained is fixed for each class. In this study, we propose a method to control the capacity of the number of core data according to the difficulty of each class, and a continual learning method CC-replay using this method. We also propose a continuous learning method, cc-replay, which uses the calculated number of core data per class. In our experiments, we evaluate the performance of cc-replay based on accuracy and training time on several benchmarks, and discuss the behavior of cc-replay when the size of the dataset and the similarity between classes are changed.