Robust Noisy-Label Learning: A Novel Class-Specific Autoencoder-Guided Sample Selection Framework

Abstract

Deep learning has been extensively applied to sample selection. However, the performance of deep learning-based sample selection methods can be substantially degraded when training data are contaminated by noisy labels, a common issue in practical industrial settings. To address this challenge, this paper proposes Class-Specific Autoencoder-Guided Sample Selection (CASS), a robust learning framework for sample selection under noisy-label conditions. In the CASS framework, a Transformer encoder is first pretrained through contrastive learning to acquire noise-resistant feature representations with enhanced capacity for modeling global feature dependencies. A class-specific autoencoder-assisted sample selection module is then introduced to characterize nonlinear feature structures and identify relatively reliable samples based on reconstruction-driven structural consistency. Finally, a center-based discriminative regularization term is incorporated into the loss function of downstream supervised training to enhance intra-class compactness and strengthen feature discriminability. By progressively coupling contrastive pretraining, class-specific autoencoder-assisted sample selection, and discriminative supervised training, CASS improves both representation learning and robustness against label noise. Experimental results on CIFAR-10 dataset with noisy labels demonstrate the effectiveness and robustness of the proposed framework.