CMCL: Clustering-based Memory Management for Continual Learning
Jiae Yoon, Hyuk Lim · 2022
Continual learning (CL) is an incremental learning method to accumulate and refine knowledge over time by continually processing datasets belonging to new tasks. While learning a new task, CL may not retain essential information on previous tasks, which is known as catastrophic forgetting (CF). The CF of information about previous tasks is a challenging problem to overcome for CL. We consider a memory-based approach to combine the previous and new data for CL and propose a memory management method using unsupervised clustering to mitigate the CF. The proposed method generates a set of clusters for the combined datasets by an unsupervised clustering and stores the most representative data belonging to each cluster in memory. The number of clusters is determined by the unsupervised clustering depending on the features of the combined dataset rather than the number of tasks. Further, an experiment was performed to compare the proposed method with existing methods that store a certain amount of data for each task in the memory. The experiment results indicate that the proposed method outperformed the existing methods.