Reproduce “Active Negative Loss: A Robust Framework for Learning with Noisy Labels” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xichen Ye, Yifan Wu, Yiqi Wang, Xiaoqiang Li, Weizhong Zhang, Yifan Chen
This paper introduces the Active Negative Loss (ANL) framework, which employs Normalized Negative Loss Functions (NNLFs) to enhance robustness and learning efficiency when training deep neural networks with noisy labels. The framework, particularly when combined with entropy-based regularization (ANL*), demonstrates superior performance and fitting ability across various synthetic and real-world noise scenarios, including image classification and segmentation tasks.
Key insights
- NNLFs are symmetric and noise-tolerant under both symmetric and asymmetric label noise, providing a theoretically sound foundation for robust learning.
- Unlike MAE, NNLFs assign higher gradients to classes with lower predicted probabilities and to samples that are better learned, enabling the model to effectively focus on clean, well-memorized examples.
- Entropy-based regularization effectively counteracts label imbalance in scenarios with high asymmetric noise, further boosting the robustness of the ANL framework.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.