Open-set Label Noise Can Improve Robustness Against Inherent Label Noise

TL;DR

This paper demonstrates that open-set noisy labels can enhance robustness against inherent label noise and introduces a regularization method using dynamic open-set samples, leading to improved model stability and OOD detection.

Contribution

It reveals the non-toxic and beneficial role of open-set noisy labels and proposes ODNL, a regularization technique that improves robustness and OOD detection in noisy label scenarios.

Findings

Open-set noisy labels can improve robustness against inherent label noise.

The proposed ODNL method enhances existing robust algorithms.

ODNL significantly improves Out-of-Distribution detection performance.

Abstract

Learning with noisy labels is a practically challenging problem in weakly supervised learning. In the existing literature, open-set noises are always considered to be poisonous for generalization, similar to closed-set noises. In this paper, we empirically show that open-set noisy labels can be non-toxic and even benefit the robustness against inherent noisy labels. Inspired by the observations, we propose a simple yet effective regularization by introducing Open-set samples with Dynamic Noisy Labels (ODNL) into training. With ODNL, the extra capacity of the neural network can be largely consumed in a way that does not interfere with learning patterns from clean data. Through the lens of SGD noise, we show that the noises induced by our method are random-direction, conflict-free and biased, which may help the model converge to a flat minimum with superior stability and enforce the model to produce conservative predictions on Out-of-Distribution instances. Extensive experimental results on benchmark datasets with various types of noisy labels demonstrate that the proposed method not only enhances the performance of many existing robust algorithms but also achieves significant improvement on Out-of-Distribution detection tasks even in the label noise setting.