Reproduce “BadLabel: A Robust Perspective on Evaluating and Enhancing Label-noise Learning” on OpenResearch

About this paper

Jingfeng Zhang, Bo Song, Haohan Wang, Bo Han, Tongliang Liu, Lei Liu, Masashi Sugiyama

The paper introduces BadLabel, a novel and challenging type of adversarial label noise designed to make clean and noisy labels indistinguishable by loss values, and proposes Robust DivideMix, an algorithm that leverages adversarial perturbations to successfully mitigate BadLabel's impact. Robust DivideMix achieves superior performance under BadLabel noise while maintaining competitive results on conventional and real-world noisy datasets.

Key insights

• Existing state-of-the-art LNL algorithms are highly vulnerable to adversarially crafted label noise (BadLabel) that makes the loss values of clean and noisy labels indistinguishable.
• Despite having similar initial loss values under BadLabel, noisy labels exhibit greater susceptibility to adversarial perturbations, leading to a significant increase in their loss compared to clean labels.
• The convergence speed of a Bayesian Gaussian Mixture Model (BayesGMM) can serve as an effective indicator for the separability quality of loss distributions, enabling adaptive and robust data filtering.

Summary by alphaXiv.