Online continual learning suffers from an underfitted solution for prompt model update due to the constraint of single-epoch learning. We confront this challenge by proposing an efficient online continual learning method with the notion of neural collapse. In particular, we induce neural collapse to form a simplex equiangular tight frame (ETF) structure in the representation space so that the learned model with single epoch can better fit the streamed data by proposing preparatory data training and residual correction in the representation space. With an extensive set of empirical validations using CIFAR10/100, TinyImageNet, and ImageNet-200, we show that our proposed method outperforms state-of-the-art methods by a noticeable margin in various online continual learning scenarios, including Disjoint and Gaussian scheduled setups.
