Learning with Noisy Labels by Adaptive Gradient-Based Outlier Removal

An accurate and substantial dataset is essential for training a reliable and well-performing model. However, even manually annotated datasets contain label errors, not to mention automatically labeled ones.