Noisy Label Learning Based on Weighted Neighborhood Consistency

In the realm of deep learning applied to real-world scenarios, the existence of noisy labels is an inevitable factor that can detrimentally affect the models’ performance. Most state-of-the-art methods for learning from noisy labels rely on sample selection strategies that partition the training data into clean and noisy labeled samples. Subsequently, these noisy label samples are treated as unlabeled samples, and the empirical vicinal risk is minimized through semi-supervised learning. Therefore, accurately identifying noisy labels contributes to enhancing the performance of the model. To enhance the accuracy and stability of sample selection, this paper proposes utilizing the mean and variance of the loss sequence to identify clean samples and noisy ones. Nonetheless, sample selection is not entirely effective in eliminating noisy label samples, as a small fraction of them are inadvertently considered as clean samples. Consequently, we propose Weighted Neighborhood Consistency Regularization (WNCR), which alleviates the impact of residual noisy labels by encouraging the neural network to maintain consistency in its predictions with those of its k-nearest neighbors for each sample. Extensive experiments on multiple synthetic and real-world noisy datasets demonstrate that our method outperforms the state-of-the-art methods at multiple noise levels.