Separating Noisy Samples From Tail Classes for Long-Tailed Image Classification With Label Noise.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Existing methods for noisy label learning often assume balanced class distributions.
• These methods struggle with imbalanced datasets, failing to distinguish noisy samples from clean samples in tail classes.

Purpose of the Study:

• To address image classification with both noisy labels and long-tailed distributions.
• To develop a robust learning paradigm for imbalanced datasets with label noise.

Main Methods:

• Propose a new learning paradigm using weak and strong data augmentations to identify noisy samples.
• Introduce leave-noise-out regularization (LNOR) to mitigate the impact of identified noisy samples.
• Implement a prediction penalty based on online classwise confidence to prevent bias towards head classes.

Main Results:

• The proposed method effectively screens noisy samples by matching inferences on different data augmentations.
• Leave-noise-out regularization successfully eliminates the influence of recognized noisy samples.
• The prediction penalty reduces bias towards dominant head classes in imbalanced datasets.
• Extensive experiments on CIFAR-10, CIFAR-100, MNIST, FashionMNIST, and Clothing1M validate the method's superiority.

Conclusions:

• The proposed approach offers a significant advancement in handling noisy labels within long-tailed image classification.
• It provides a robust solution for practical scenarios with imbalanced data distributions and label noise.