Abstract:There has been significant attention devoted to the effectiveness of various domains, such as semi-supervised learning, contrastive learning, and meta-learning, in enhancing the performance of methods for noisy label learning (NLL) tasks. However, most existing methods still depend on prior assumptions regarding clean samples amidst different sources of noise (\eg, a pre-defined drop rate or a small subset of clean samples). In this paper, we propose a simple yet powerful idea called \textbf{NPN}, which revolutionizes \textbf{N}oisy label learning by integrating \textbf{P}artial label learning (PLL) and \textbf{N}egative learning (NL). Toward this goal, we initially decompose the given label space adaptively into the candidate and complementary labels, thereby establishing the conditions for PLL and NL. We propose two adaptive data-driven paradigms of label disambiguation for PLL: hard disambiguation and soft disambiguation. Furthermore, we generate reliable complementary labels using all non-candidate labels for NL to enhance model robustness through indirect supervision. To maintain label reliability during the later stage of model training, we introduce a consistency regularization term that encourages agreement between the outputs of multiple augmentations. Experiments conducted on both synthetically corrupted and real-world noisy datasets demonstrate the superiority of NPN compared to other state-of-the-art (SOTA) methods. The source code has been made available at {\color{purple}{\url{<a class="link-external link-https" href="https://github.com/NUST-Machine-Intelligence-Laboratory/NPN" rel="external noopener nofollow">this https URL</a>}}}.

What problem does this paper attempt to address?

The paper primarily focuses on addressing the problem of Noisy Label Learning (NLL), particularly on how to improve the robustness and classification performance of models in the presence of low-quality samples and noisy labels. Specifically: 1. **Research Background**: With the development of large-scale annotated datasets (such as ImageNet), supervised learning has made significant progress. However, obtaining high-quality annotated data has become increasingly difficult and time-consuming, limiting the scalability of models. As a result, weakly supervised learning (such as noisy label learning) has received widespread attention. 2. **Core Issue**: Existing noisy label learning methods mostly rely on some prior assumptions, such as predefined noise rates or a small portion of clean samples. These methods face challenges in handling noisy labels, especially when the noise rate is high, as models tend to overfit the noisy labels, thereby affecting the final classification performance. 3. **Solution**: The paper proposes a new method called NPN, which combines Partial Label Learning (PLL) and Negative Learning (NL) to decompose the given label space in an adaptive, data-driven manner and enhance the model's robustness using both direct and indirect supervision information. Specifically, it includes: - **Partial Label Learning (PLL)**: Each training sample is assigned a candidate label set that includes the true label. - **Negative Learning (NL)**: All non-candidate labels are used as complementary labels to indirectly guide the network that "the input image does not belong to these complementary labels." 4. **Experimental Validation**: Experiments on synthetic noisy datasets (CIFAR100N) and three real-world datasets (Web-Aircraft, Web-Car, Web-Bird) demonstrate that NPN outperforms existing state-of-the-art methods under different types of noise conditions. In summary, this paper aims to overcome the limitations of traditional noisy label learning methods and improve the classification accuracy and robustness of models in noisy environments by introducing a new framework, NPN.

Adaptive Integration of Partial Label Learning and Negative Learning for Enhanced Noisy Label Learning

ALIM: Adjusting Label Importance Mechanism for Noisy Partial Label Learning

Partial Label Learning with Noisy Side Information

Learning from Noisy Labels with Decoupled Meta Label Purifier

Noisy Label Tolerance: A New Perspective of Partial Multi-Label Learning.

Towards Robust Learning with Noisy and Pseudo Labels for Text Classification

Pseudo-labelling meets Label Smoothing for Noisy Partial Label Learning

LNL+K: Enhancing Learning with Noisy Labels Through Noise Source Knowledge Integration

PNP: Robust Learning from Noisy Labels by Probabilistic Noise Prediction

Learning With Noisy Labels Over Imbalanced Subpopulations

BPT-PLR: A Balanced Partitioning and Training Framework with Pseudo-Label Relaxed Contrastive Loss for Noisy Label Learning

Fine-Grained Classification with Noisy Labels

BadLabel: A Robust Perspective on Evaluating and Enhancing Label-noise Learning

Network Cooperation with Progressive Disambiguation for Partial Label Learning

Negative Label and Noise Information Guided Disambiguation for Partial Multi-Label Learning

An Empirical Study on Noisy Label Learning for Program Understanding

Learning with Noisy Labels Via Self-supervised Adversarial Noisy Masking

IRNet: Iterative Refinement Network for Noisy Partial Label Learning

Active Negative Loss: A Robust Framework for Learning with Noisy Labels