This paper proposes a method called Equi-Angular Representation Learning (EARL) for online continual learning (OCL), which addresses the challenge of learning from a continuous stream of data without multiple epochs of training. The method leverages the neural collapse phenomenon, where the activations of the last layer and classifier vectors form a simplex equiangular tight frame (ETF) structure in a balanced dataset. To induce neural collapse in online OCL, the authors propose two key components: preparatory data training and residual correction. Preparatory data training involves generating synthetic data that is different from existing classes to prevent new classes from being biased towards existing classes. This helps in accelerating convergence towards the ETF structure. Residual correction is used during inference to compensate for the fact that the model may not have fully converged to the ETF structure due to the continuous stream of new data. By storing residuals between the last layer activation and the classifier vector during training, the model can correct the inference output during inference to improve anytime inference accuracy. The authors evaluate their method on several datasets including CIFAR-10, CIFAR-100, TinyImageNet, ImageNet-200, and ImageNet-1K. They show that their method outperforms state-of-the-art methods in various online continual learning scenarios, particularly in terms of anytime inference accuracy. The method is effective in both disjoint and Gaussian scheduled continuous data setups. The proposed method is implemented using a fixed ETF classifier and a dot regression loss function. The results demonstrate that EARL achieves significant improvements in performance, especially in scenarios where the model needs to make predictions at any point during training. The method is also efficient in terms of memory usage and computational resources.This paper proposes a method called Equi-Angular Representation Learning (EARL) for online continual learning (OCL), which addresses the challenge of learning from a continuous stream of data without multiple epochs of training. The method leverages the neural collapse phenomenon, where the activations of the last layer and classifier vectors form a simplex equiangular tight frame (ETF) structure in a balanced dataset. To induce neural collapse in online OCL, the authors propose two key components: preparatory data training and residual correction. Preparatory data training involves generating synthetic data that is different from existing classes to prevent new classes from being biased towards existing classes. This helps in accelerating convergence towards the ETF structure. Residual correction is used during inference to compensate for the fact that the model may not have fully converged to the ETF structure due to the continuous stream of new data. By storing residuals between the last layer activation and the classifier vector during training, the model can correct the inference output during inference to improve anytime inference accuracy. The authors evaluate their method on several datasets including CIFAR-10, CIFAR-100, TinyImageNet, ImageNet-200, and ImageNet-1K. They show that their method outperforms state-of-the-art methods in various online continual learning scenarios, particularly in terms of anytime inference accuracy. The method is effective in both disjoint and Gaussian scheduled continuous data setups. The proposed method is implemented using a fixed ETF classifier and a dot regression loss function. The results demonstrate that EARL achieves significant improvements in performance, especially in scenarios where the model needs to make predictions at any point during training. The method is also efficient in terms of memory usage and computational resources.