Robust Hyperspherical Classification with Dynamic Assignment Based on RepViT

Existing lightweight classification models suffer from low accuracy and overconfidence when the data exhibits a long-tailed distribution. To address the above problems, based on RepViT network, a robust dynamic hypersphere classification method is designed in this paper. First, an adaptive label-to-prototype assignment method is proposed, which can fully utilize the feature information and dynamically optimize the assignment during the training process. Second, a robust enhancement method is designed by combining kinds of techniques including regularization, classifiers, and optimization techniques to reduce the overconfidence in false predictions. The method in this paper improves accuracy by 3% on both standard datasets such as Cifar-10 and its long-tailed variants. It also reduces the accuracy loss by 15% on long-tailed datasets compared to the rest of the models.