HEALON: Progressive hard sample attenuation for learning with noisy labels

Abstract

Noisy labels are often present in large, accessible datasets. Learning with noisy labels can degrade the generalization performance of DNNs. While Semi-Supervised Learning (SSL) approaches have shown promise by predicting pseudo-labels to correct noisy labels, we find and demonstrate a fundamental limitation of SSL-based label correction methods: hard samples near decision boundaries significantly weaken the memorization effect that these methods rely on. This leads to erroneous pseudo-labels and creates a negative feedback loop where models gradually memorize these errors, further degrading performance. To overcome the issue, inspired by AdaBoost and building on these insights, we propose HEALON (Hard samplE Adaptive Labeling with Optimal reweighting for Noisy labels), a novel framework for learning with noisy labels that effectively addresses the hard sample challenge through an optimal weighting strategy that balances their influence during training. The framework primarily consists of two steps: Weighted Implicit Ensemble (WIE) and Weight Optimization for HArd Sample (WOHAS). WIE combines the Adaboost strategy with multiple sets of weight distributions obtained from WOHAS to train a single model, allowing the model to converge to multiple local optima along its optimization path and predict pseudo-labels. A sample difficulty minimization method is then designed to aggregate the predicted labels, generating near-global optimal pseudo-labels for each noisy sample, followed by a multiple “snapshot” weights strategy to minimize computational cost. WOHAS quantifies sample difficulty using information entropy and derives optimal weights to WIE via a cumulative sample difficulty strategy, balancing the impact of hard samples while preserving the memorization effect. Extensive experiments on benchmark datasets demonstrate that our approach significantly outperforms state-of-the-art methods in both pseudo-label correction accuracy and overall classification performance.