Long-Tailed Noisy Labels Learning via Cognitively-Inspired External Guidance
Abstract
Learning with long-tailed noisy labels presents a fundamental challenge for classification, as the combined effects of class imbalance and label corruption distort the feature space. These distortions lead to two forms of inseparability: i) semantic confusion, where mislabeled samples resemble semantically similar classes, and ii) structural degradation, where clean tail samples are misclassified as noise. Existing methods typically depend on internal visual features, which are themselves biased and corrupted, making reliable sample selection intrinsically difficult. In this paper, we propose a paradigm shift from feature-based heuristics to external knowledge-guided verification. Drawing inspiration from the human coarse-to-fine cognitive process, we introduce CoEG (Cognitively-Inspired External Guidance), a novel framework that synthesizes holistic perception from vision-language models and analytical disambiguation from large language models. This dual-system design yields robust cognitive confidence for noise verification, enabling accurate and interpretable sample selection. To further enhance representation quality, CoEG employs relational knowledge distillation to transfer structured topological relations from external models to correct the student's distorted feature manifold. Extensive experiments on both synthetic and real-world long-tailed noisy datasets demonstrate the effectiveness of CoEG.