The paper "Learning Equi-angular Representations for Online Continual Learning" addresses the challenge of online continual learning, where models must adapt to new data streams without forgetting past tasks. The authors propose a method called Equi-Angular Representation Learning (EARL) to induce neural collapse, a phenomenon where the activations of the last layer and classifier vectors form a simplex equiangular tight frame (ETF) structure. This structure ensures that all pairwise angles between classes are equal and maximally widened, improving model performance. To overcome the limitations of single-epoch training in online continual learning, EARL introduces two key components: preparatory data training and residual correction. Preparatory data training involves synthesizing out-of-distribution (OOD) samples from existing classes to prevent new classes from being biased towards existing ones. This helps mitigate the bias problem, which hinders the convergence of features to the ETF structure. Residual correction addresses the remaining discrepancy between the model's predictions and the ETF classifier during inference, ensuring better anytime inference accuracy. The authors evaluate EARL on various datasets and setups, including CIFAR-10, CIFAR-100, TinyImageNet, ImageNet-200, and ImageNet-1K. Empirical results show that EARL outperforms state-of-the-art methods in terms of anytime inference performance, achieving significant gains in area under the curve accuracy (AUC) and last accuracy. The paper also includes an ablation study to validate the effectiveness of each component and discusses the limitations and future directions of the proposed approach.The paper "Learning Equi-angular Representations for Online Continual Learning" addresses the challenge of online continual learning, where models must adapt to new data streams without forgetting past tasks. The authors propose a method called Equi-Angular Representation Learning (EARL) to induce neural collapse, a phenomenon where the activations of the last layer and classifier vectors form a simplex equiangular tight frame (ETF) structure. This structure ensures that all pairwise angles between classes are equal and maximally widened, improving model performance. To overcome the limitations of single-epoch training in online continual learning, EARL introduces two key components: preparatory data training and residual correction. Preparatory data training involves synthesizing out-of-distribution (OOD) samples from existing classes to prevent new classes from being biased towards existing ones. This helps mitigate the bias problem, which hinders the convergence of features to the ETF structure. Residual correction addresses the remaining discrepancy between the model's predictions and the ETF classifier during inference, ensuring better anytime inference accuracy. The authors evaluate EARL on various datasets and setups, including CIFAR-10, CIFAR-100, TinyImageNet, ImageNet-200, and ImageNet-1K. Empirical results show that EARL outperforms state-of-the-art methods in terms of anytime inference performance, achieving significant gains in area under the curve accuracy (AUC) and last accuracy. The paper also includes an ablation study to validate the effectiveness of each component and discusses the limitations and future directions of the proposed approach.