# Robust and On-the-fly Dataset Denoising for Image Classification

arXiv: https://arxiv.org/abs/2003.10647 · HTML: https://mlsanity.com/paper/2003.10647/

## Key facts

- Cited by 6 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 24 Mar 2020.

## Built on by (most cited first, 6 of 6)

- [Open-World Semi-Supervised Learning](https://mlsanity.com/paper/2102.03526/) (2021)
- [Contrast to Divide: Self-Supervised Pre-Training for Learning with Noisy Labels](https://mlsanity.com/paper/2103.13646/) (2021)
- [Learning to Combat Noisy Labels via Classification Margins](https://mlsanity.com/paper/2102.00751/) (2021)
- [Leveraging Unlabeled Data to Track Memorization](https://mlsanity.com/paper/2212.04461/) (2022)
- [FPR Estimation for Fraud Detection in the Presence of Class-Conditional Label Noise](https://mlsanity.com/paper/2308.02695/) (2023)
- [Label Denoising through Cross-Model Agreement](https://mlsanity.com/paper/2308.13976/) (2023)

## Abstract (arXiv)

Memorization in over-parameterized neural networks could severely hurt generalization in the presence of mislabeled examples. However, mislabeled examples are hard to avoid in extremely large datasets collected with weak supervision. We address this problem by reasoning counterfactually about the loss distribution of examples with uniform random labels had they were trained with the real examples, and use this information to remove noisy examples from the training set. First, we observe that examples with uniform random labels have higher losses when trained with stochastic gradient descent under large learning rates. Then, we propose to model the loss distribution of the counterfactual examples using only the network parameters, which is able to model such examples with remarkable success. Finally, we propose to remove examples whose loss exceeds a certain quantile of the modeled loss distribution. This leads to On-the-fly Data Denoising (ODD), a simple yet effective algorithm that is robust to mislabeled examples, while introducing almost zero computational overhead compared to standard training. ODD is able to achieve state-of-the-art results on a wide range of datasets including real-world ones such as WebVision and Clothing1M.