What question did this study set out to answer?

To develop a simple and effective method for detecting noisy training labels in deep neural networks without requiring human intervention or specialized noise-robust architectures.

October 1, 2019

O2U-Net: A Simple Noisy Label Detection Approach for Deep Neural Networks

View Full Paper

Ask AI

Bookmark

Authors

JHJinchi HuangAlibaba Group (United States)LQLie QuAlibaba Group (China)RJRongfei JiaPeking University

Discussion

Member takes

Implication

Machine learning study reveals cyclic overfitting and underfitting identifies corrupted training labels in deep neural networks, indicating a simple alternative to complex noise-robust loss functions.

Key Points

To develop a simple and effective method for detecting noisy training labels in deep neural networks without requiring human intervention or specialized noise-robust architectures.
Adjusted model hyperparameters cyclically to transition the network between states of overfitting and underfitting across training iterations.
Recorded individual sample losses across cycles, identifying mislabeled instances based on higher normalized average loss values.
Evaluated label detection performance across multiple benchmark datasets under diverse experimental noise configurations.
Demonstrated that samples displaying persistently high normalized average loss across cyclical phases reliably indicate noisy labels.
Achieved state-of-the-art noisy label detection performance across multiple datasets while maintaining compatibility with active learning frameworks.

Cite This Study

Huang et al. (2019) studied this question.

synapsesocial.com/papers/6a0ebba4950456576347b656 https://doi.org/10.1109/iccv.2019.00342

View Full Paper

Ask AI

Bookmark