Entropy-based Sample Selection for Online Continual Learning
Felix Wiewel, Bin Yang
Abstract
Deep neural networks (DNNs) suffer from catastrophic forgetting, a rapid decrease in performance when trained on a sequence of tasks where only data of the most recent task is available. Most previous research has focused on the case where all data of a task is available simultaneously and boundaries between tasks are known. In this paper, we focus on the online setting where data arrives one-by-one or in small batches ordered by tasks and task boundaries are unknown. Avoiding catastrophic forgetting in such a setting is of great interest since it would allow DNNs to accumulate knowledge without the need to store all previously seen data even if task boundaries are unknown. For this, we propose a novel rehearsal algorithm for online continual learning that is derived from basic concepts of information theory. We demonstrate on commonly used data sets that our method can avoid catastrophic forgetting, achieve competitive results when compared with the current state-of-the-art and even outperform it in most cases.