Enhanced Experience Replay for Class Incremental Continual Learning

Continual learning aims to construct a machine model that learns multiple tasks sequentially. However, it may face the significant challenge of catastrophic forgetting, where the model minimizes the loss on the current data and forgets the previously learned tasks. A common approach to mitigate catastrophic forgetting is replay-based continual learning, which stores data points from previous tasks in a buffer and revisits them periodically. However, not all data points in a task contribute the same significance for learning. Hence, coreset selection is crucial for continual learning with imbalanced and noisy datasets. In this paper, we introduce Enhanced Experience Replay (EER), a simple yet effective method that selects the most representative and informative coreset for training at each iteration. EER not only optimizes the model's adaptability to the current dataset but also prioritizes samples exhibiting a high affinity with previous tasks. Evaluated on CIFAR-10 and CIFAR-100 datasets, our coreset selection mechanism significantly enhances task adaptability and prevent catastrophic forgetting. The proposed method achieves state-of-the-art performance on the two benchmark datasets.