This repository contains the code for the paper

Using Trusted Data to Train Deep Networks on Labels Corrupted by Severe Noise (NeurIPS 2018).

The code requires Python 3+, PyTorch [0.3, 0.4), and TensorFlow (for loading MNIST).

The Gold Loss Correction (GLC) is a semi-verified method for label noise robustness in deep learning classifiers. Using a small set of data with trusted labels, we estimate parameters of the label noise, which we then use to train a corrected classifier on the noisy labels. We observe large gains in performance over prior work, with a subset of results shown below. Please consult the paper for the full results and method descriptions.

To obtain accuracies, run the following scripts.

Non-CIFAR: python <dataset>_experiments_pytorch.py --method $1 --corruption_type $2

CIFAR: python train_<method>.py --gold_fraction $1 --corruption_prob $2 --corruption_type $3

Change 'dataset', 'method', and the command line arguments to specify the experiment to be run. The non-CIFAR scripts return percent accuracies for all gold fractions and corruption probabilities, while the CIFAR scripts only give one accuracy value at a time. Area under the error curve can be obtained by running numpy.trapz on the list of percent errors for corruption probabilities from 0.1 to 1.0 inclusive.

If you find this useful in your research, please consider citing:

@article{hendrycks2018glc,
  title={Using Trusted Data to Train Deep Networks on Labels Corrupted by Severe Noise},
  author={Hendrycks, Dan and Mazeika, Mantas and Wilson, Duncan and Gimpel, Kevin},
  journal={Advances in Neural Information Processing Systems},
  year={2018}
}