Distribution-Aware Robust Learning from Long-Tailed Data with Noisy Labels

TL;DR

DaSC presents a robust approach to learning from long-tailed data with noisy labels by introducing distribution-aware class centroid estimation (DaCC) and confidence-aware contrastive learning. DaCC weights samples by model predictions (with temperature scaling) to form centroids using all data, while a Gaussian Mixture Model separates clean from noisy data. The training framework splits samples into high- and low-confidence groups, applying Semi-supervised Balanced Contrastive Loss to the high-confidence set and Mixup-enhanced Instance Discrimination Loss to the low-confidence set, integrated within a semi-supervised learning paradigm. Empirical results on long-tailed CIFAR variants and real-world noisy-label datasets demonstrate state-of-the-art performance and robustness across varying noise types and imbalance ratios.

Abstract

Deep neural networks have demonstrated remarkable advancements in various fields using large, well-annotated datasets. However, real-world data often exhibit long-tailed distributions and label noise, significantly degrading generalization performance. Recent studies addressing these issues have focused on noisy sample selection methods that estimate the centroid of each class based on high-confidence samples within each target class. The performance of these methods is limited because they use only the training samples within each class for class centroid estimation, making the quality of centroids susceptible to long-tailed distributions and noisy labels. In this study, we present a robust training framework called Distribution-aware Sample Selection and Contrastive Learning (DaSC). Specifically, DaSC introduces a Distribution-aware Class Centroid Estimation (DaCC) to generate enhanced class centroids. DaCC performs weighted averaging of the features from all samples, with weights determined based on model predictions. Additionally, we propose a confidence-aware contrastive learning strategy to obtain balanced and robust representations. The training samples are categorized into high-confidence and low-confidence samples. Our method then applies Semi-supervised Balanced Contrastive Loss (SBCL) using high-confidence samples, leveraging reliable label information to mitigate class bias. For the low-confidence samples, our method computes Mixup-enhanced Instance Discrimination Loss (MIDL) to improve their representations in a self-supervised manner. Our experimental results on CIFAR and real-world noisy-label datasets demonstrate the superior performance of the proposed DaSC compared to previous approaches.

Figures (11)

• Figure 1: (a) Class distribution on long-tailed CIFAR-10 with symmetric noise of 0.6. (b) and (c) show the illustration of previous class centroid estimation and our distribution-aware class centroid estimation, respectively. In the previous methods, samples with low confidence, depicted as grey dots, are excluded from the class centroid estimation. In contrast, our approach leverages all samples for centroid estimation, using their prediction values as weights. The colors of the samples indicate the prediction values.
• Figure 2: Illustration of the proposed DaSC framework: Given a dataset, the DaSC detects correctly labeled samples using class centroids obtained from DaCC. Subsequently, the model is trained under a semi-supervised learning (SSL) framework, where confidence-aware contrastive learning is employed to enhance the representations. $\texttt{Cat}$ and $\texttt{sg}$ denotes concatenation and stop-gradient operations, respectively.
• Figure 3: Accuracy of pseudo-labels for different sample groups.
• Figure 4: Ablation study on the long-tailed version of CIFAR.
• Figure 4: Performance of the DaSC compared with the baseline method for different class sets including 'Many', 'Medium', and 'Few'.