Balancing Explanations and Adaptation in Offline Continual Learning Systems Using Active Augmented Reply
Md Abdullah Al Forhad, Weishi Shi · 2024
Offline continual learning (CL) aims to learn multiple task datasets sequentially, which leads to the well-known problem of catastrophic forgetting. Buffer or memory-based methods mitigate catastrophic forgetting by storing subsets of task experiences in memory. This approach has demonstrated state-of-the-art performance in existing works. However, most research focuses on improving the overall model while using uniform selection for storing experiences in the buffer. Although uniform sampling has shown competitive performance in various studies, systematic sampling should yield greater overall gains. There is limited research focus in this domain. The active learning community has demonstrated the benefits of active sampling over uniform sampling in both traditional machine learning and deep learning settings. However, active learning in a continual learning setting is not as straightforward as traditional active learning because sudden distribution changes are common in CL when moving from one dataset to another, which can cause the sampling method to be biased toward one particular dataset. This inspires us to design an active sampling method for the memory buffer in a CL setting. To mitigate the effect of sudden distribution change while maximizing buffer utilization to reduce the forgetting effect, we design a dual memory buffer setting with active and augmented sampling, incorporating some allowed randomness. We conduct thorough experiments and ablation studies, which show that dual memory-based active and augmented sampling enhances memory utilization while reducing the forgetting effect. Our results support our claim and provide competitive performance compared to existing baselines on multiple datasets. This will inspire the research community to focus more on this direction.