Dual-Criterion Sample Selection for Noisy Labels: Integrating Neighborhood Prediction Divergence and Loss Values

Learning with noisy labels is a fundamental challenge in deep learning, as mislabeled samples can lead to severe overfitting and performance degradation. A widely used strategy for learning with noisy labels is sample selection based on the small-loss criterion, which assumes low-loss samples are more likely clean. However, this assumption fails when noisy samples also exhibit low losses—due to memorization—causing error accumulation during training. We propose a novel sample selection strategy that jointly considers Neighborhood Prediction Divergence (NPD) and loss values to more reliably identify clean samples. NPD captures prediction divergence within a sample’s semantic neighborhood, providing a complementary signal to loss for distinguishing noisy labels. Only samples with low divergence and low loss are selected as clean. Experiments on CIFAR-10/100, WebVision, and Food-101N show that our method significantly improves clean-sample F1-scores and consistently boosts the performance of state-of-the-art noisy-label learning frameworks.