Using samples with label noise for robust continual learning

Abstract

Recent studies have shown that effectively leveraging samples with label noise can enhance model robustness by uncovering more reliable feature patterns. While existing methods, such as label correction methods and loss correction techniques, have demonstrated success in utilizing noisy labels, they assume that noisy and clean samples (samples with correct annotations) share the same label space.However, this assumption does not hold in continual machine learning, where new categories and tasks emerge over time, leading to label shift problems that are specific to this setting. As a result, existing methods may struggle to accurately estimate the ground truth labels for noisy samples in such dynamic environments, potentially exacerbating label noise and further degrading performance. To address this critical gap, we propose a Shift-Adaptive Noise Utilization (SANU) method, designed to transform samples with label noise into usable samples for continual learning. SANU introduces a novel source detection mechanism that identifies the appropriate label space for noisy samples, leveraging a meta-knowledge representation module to improve the generalization of the detection process. By re-annotating noisy samples through label guessing and label generation strategies, SANU adapts to label shifts, turning noisy data into useful inputs for training. Experimental results across three continual learning datasets demonstrate that SANU effectively mitigates the label shift problem, significantly enhancing model performance by utilizing re-annotated samples with label noise.