Complementary-Label Learning

This repository gives the implementation for complementary-label learning from the ICML 2019 paper [1], the ECCV 2018 paper [2], and the NeurIPS 2017 paper [3].

Requirements

• Python 3.6
• numpy 1.14
• PyTorch 1.1
• torchvision 0.2

Demo

The following demo will show the results with the MNIST dataset. After running the code, you should see a text file with the results saved in the same directory. The results will have three columns: epoch number, training accuracy, and test accuracy.

python demo.py -h

Methods and models

In demo.py, specify the method argument to choose one of the 5 methods available:

• ga: Gradient ascent version (Algorithm 1) in [1].
• nn: Non-negative risk estimator with the max operator in [1].
• free: Assumption-free risk estimator based on Theorem 1 in [1].
• forward: Forward correction method in [2].
• pc: Pairwise comparison with sigmoid loss in [3].

Specify the model argument:

• linear: Linear model
• mlp: Multi-layer perceptron with one hidden layer (500 units)

Reference

1. T. Ishida, G. Niu, A. K. Menon, and M. Sugiyama.
   Complementary-label learning for arbitrary losses and models.
   In ICML 2019.
   [paper]
2. Yu, X., Liu, T., Gong, M., and Tao, D.
   Learning with biased complementary labels.
   In ECCV 2018.
   [paper]
3. T. Ishida, G. Niu, W. Hu, and M. Sugiyama.
   Learning from complementary labels.
   In NeurIPS 2017.
   [paper]

If you have any further questions, please feel free to send an e-mail to: ishida at ms.k.u-tokyo.ac.jp.