Classifying Long-tailed and Label-noise Data via Disentangling and Unlearning

TL;DR

This work tackles the intertwined challenges of long-tailed distributions and label noise by introducing tail-to-head ($T2H$) noise as a practical, unidirectional degradation. It proposes DULL, a disentangle-and-unlearn framework comprising Inner-Feature Disentangling (IFD) and Inner-Feature Partial Unlearning (IFPU), augmented by Adaptive Multi-Labeling and Head-to-Tail Knowledge Transfer to robustly learn under $T2H$ noise. The method uses a learnable channel-class correlation matrix $G$ to disentangle features and selectively erase erroneous feature regions, guided by multi-label corrections and soft-label mixup. Empirical results on simulated CIFAR datasets and real-world benchmarks (Clothing1M, WebVision-50) show that DULL consistently outperforms existing LT, NL, and LTNL approaches, demonstrating strong robustness to challenging real-world noise and imbalanced data. The work advances practical strategies for learning under joint long-tailed and noisy-label conditions and highlights areas for scalability improvements in feature-disentangling approaches as class counts grow.

Abstract

In real-world datasets, the challenges of long-tailed distributions and noisy labels often coexist, posing obstacles to the model training and performance. Existing studies on long-tailed noisy label learning (LTNLL) typically assume that the generation of noisy labels is independent of the long-tailed distribution, which may not be true from a practical perspective. In real-world situaiton, we observe that the tail class samples are more likely to be mislabeled as head, exacerbating the original degree of imbalance. We call this phenomenon as ``tail-to-head (T2H)'' noise. T2H noise severely degrades model performance by polluting the head classes and forcing the model to learn the tail samples as head. To address this challenge, we investigate the dynamic misleading process of the nosiy labels and propose a novel method called Disentangling and Unlearning for Long-tailed and Label-noisy data (DULL). It first employs the Inner-Feature Disentangling (IFD) to disentangle feature internally. Based on this, the Inner-Feature Partial Unlearning (IFPU) is then applied to weaken and unlearn incorrect feature regions correlated to wrong classes. This method prevents the model from being misled by noisy labels, enhancing the model's robustness against noise. To provide a controlled experimental environment, we further propose a new noise addition algorithm to simulate T2H noise. Extensive experiments on both simulated and real-world datasets demonstrate the effectiveness of our proposed method.

Figures (5)

• Figure 1: A case study of tail-to-head (T2H) noisy and long-tailed distribution of CIFAR-100 with an original imbalance factor of 10.
• Figure 2: (a) In epoch $k$, the model extracts salient duck features (blue regions), producing higher prediction in duck class. However, the prediction and the noisy label 'dog' result in a high loss, which adjusts the model to reinforce and output salient dog feature by back-propagation. In epoch $k+1$, the updated model outputs more salient dog features (red regions) and less salient duck features for the same duck sample. This illustrates how the model is misled by noisy labels, leading to a degradation in classification performance. (b) Illustration of the core of IFPU mechanism, show how it selectively unlearns incorrect feature regions associations to wrong classes, preventing model's wrong reinforcement, thereby enhancing robustness against noisy data.
• Figure 3: Illustration of our proposed DULL method. The left panel shows the Inner-Feature Disentangling (IFD) mechanism, which aims to separate feature channels into independent regions, ensuring that deactivating one feature region does not impact others. Based on the disentangled features, the Inner-Feature Partial Unlearning (IFPU) mechanism, illustrated in the right panel, unlearns incorrect feature regions associated with wrong classes, thereby preventing the model from reinforcing incorrect information.
• Figure 4: Hyperparameter sensitivity analysis of $\beta$ in IKD on CIFAR10 with an original IF set to 10 and a simulated T2H noise ratio of 40%.
• Figure 5: Hit rate of corrected labels for FixMatch and multi-label mechanism of DULL under different simulated T2H noise ratios on the long-tailed CIFAR-100 dataset with an original IF set to 10.