Learning with Open-world Noisy Data via Class-independent Margin in Dual Representation Space

TL;DR

This work tackles Learning with Open-world Noisy Data (LOND), where training data contain both closed-set mislabels and open-set noise. It introduces a dual-representation framework with a prototype-space projection network and a class-independent One-Vs-All (OVA) network to jointly learn discriminative representations and detect open-set samples. By combining bi-level contrastive learning, consistency regularization, and a class-independent margin for sample identification, the method robustly distinguishes clean data, closed-set noise, and open-set noise. Experiments on CIFAR80N, CIFAR100N, and Web datasets demonstrate state-of-the-art performance, with average accuracy improvements of around 4.55% and AUROC gains of about 6.17% on CIFAR80N, highlighting the value of dual-space learning for robust classification and open-set detection in noisy, open-world settings.

Abstract

Learning with Noisy Labels (LNL) aims to improve the model generalization when facing data with noisy labels, and existing methods generally assume that noisy labels come from known classes, called closed-set noise. However, in real-world scenarios, noisy labels from similar unknown classes, i.e., open-set noise, may occur during the training and inference stage. Such open-world noisy labels may significantly impact the performance of LNL methods. In this study, we propose a novel dual-space joint learning method to robustly handle the open-world noise. To mitigate model overfitting on closed-set and open-set noises, a dual representation space is constructed by two networks. One is a projection network that learns shared representations in the prototype space, while the other is a One-Vs-All (OVA) network that makes predictions using unique semantic representations in the class-independent space. Then, bi-level contrastive learning and consistency regularization are introduced in two spaces to enhance the detection capability for data with unknown classes. To benefit from the memorization effects across different types of samples, class-independent margin criteria are designed for sample identification, which selects clean samples, weights closed-set noise, and filters open-set noise effectively. Extensive experiments demonstrate that our method outperforms the state-of-the-art methods and achieves an average accuracy improvement of 4.55\% and an AUROC improvement of 6.17\% on CIFAR80N.

Figures (6)

• Figure 1: Illustration of the LOND setup and the effect of open-set noise. Left: In addition to closed-set noise, open-set noise is also present in the training and testing stage. Right: Open-set noise (w/ Open) significantly degrades performance on CIFAR80N with 80% symmetric noise.
• Figure 2: The overall framework of our proposed method. It uses projection and OVA networks to jointly learn in dual representation space, where bi-level contrastive learning and consistency regularization are introduced to enhance the detection of open-set noise. Then, class-independent margin criteria are used for sample identification. It uses the neighbor margin to select class-balanced clean samples $D_{clean}$, weighted closed-set noise $D_{close}$, and the negative margin to filter open-set noise $D_{open}$. Different losses are applied to these sample sets to obtain a classifier for known classes and a detector for open-set noise.
• Figure 3: The sensitivity of hyper-parameters $\alpha_{ID}$, $\alpha_{OOD}$, $\lambda_{Con}$, and $\lambda_{BCL}$, where "Avg." denotes the average of accuracy and AUROC of three cases (Sym-20%, Sym-80%, and Asym-40%) on CIFAR80N.
• Figure 4: The sensitivity of $K$ on CIFAR80N with symmetric noise (left) and asymmetric noise (right).
• Figure 5: Comparison of our method and UNICON on CIFAR80N with 40% asymmetric noise rate.