NLC-block: Enhancing neural network training robustness with noisy label reweighting

Abstract

Noisy labels pose a major challenge in supervised learning, often undermining the reliability and generalization of deep neural networks. Addressing this issue requires mitigating the adverse impact of mislabeled samples and avoiding overly complex architectures or extended training procedures. To this end, this paper proposes the NLC block (Noisy Label Correction), a lightweight, plug-and-play module inspired by the \(\gamma \)-divergence weighting principle. Unlike traditional parameter-dependent methods, the NLC block integrates a feed-forward layer with a closed-form formula computation layer to dynamically reweight samples without introducing additional learnable parameters. This paper provides a theoretical analysis demonstrating its robustness and shows, through extensive experiments on real-world datasets, that the NLC block significantly improves model accuracy and stability under label noise. The implementation is publicly available at https://github.com/DebtVC2022/NLC-block.