Reproduce “Classifying Long-tailed and Label-noise Data via Disentangling and Unlearning” on OpenResearch

About this paper

Chen Shu, Mengke Li, Yiqun Zhang, Yang Lu, Bo Han, Yiu-ming Cheung, Hanzi Wang

The DULL framework addresses the challenge of tail-to-head (T2H) noise in long-tailed and label-noisy datasets, which systematically biases models against minority classes. It achieves improved accuracy, such as 80.43% on CIFAR-10 with 40% T2H noise and 74.12% on Clothing1M, by disentangling feature associations and selectively unlearning misleading information.

Key insights

• Tail-to-head (T2H) noise systematically biases models by forcing them to learn incorrect feature associations between true and mislabeled classes, particularly harming minority classes.
• Disentangling feature channels enables selective weakening of misleading associations from noisy labels while preserving correct feature representations.
• An "unlearning" paradigm can effectively mitigate the impact of specific, harmful noise patterns by actively adjusting model weights rather than solely relying on noisy sample detection and rejection.

Summary by alphaXiv.