Continual Learning for Long-Tailed Recognition: Bridging the Gap in Theory and Practice

The Long-Tailed Recognition (LTR) problem arises in imbalanced datasets. This paper bridges the theory-practice gap in this context, providing mathematical insights into the training dynamics of LTR scenarios by proposing a theorem stating that, under strong convexity, the learner's weights trained on the full dataset are bounded by those trained only on the Head. We extend this theorem for multiple subsets and introduce a novel perspective of using Continual Learning (CL) for LTR. We sequentially learn the Head and Tail by updating the learner's weights without forgetting the Head using CL methods. We prove that CL reduces loss compared to fine-tuning on the Tail. Our experiments on MNIST-LT and standard LTR benchmarks (CIFAR100-LT, CIFAR10-LT, and ImageNet-LT) validate our theory and demonstrate the effectiveness of CL solutions. We also show the efficacy of CL on real-world data, specifically the Caltech256 dataset, outperforming state-of-the-art classifiers. Our work unifies LTR and CL and paves the way for leveraging advances in CL to tackle the LTR challenge effectively.