Keras implementation of UU learning

This is a reproducing code for the ICLR'19 paper: On the Minimal Supervision for Training Any Binary Classifier from Only Unlabeled Data.

• loss.py has a keras implementation of the risk estimator for UU learning (see Eq.(10) in the paper) and its simplified version (see Eq.(12) in the paper).

• experiment.py is an example code of UU learning.

Datasets are MNIST preprocessed in such a way that even digits form the P class and odd digits form the N class, and CIFAR10 preprocessed in such a way that the P class is composed of 'bird', 'cat', 'deer', 'dog', 'frog' and 'horse'; the N class is composed of 'airplane', 'automobile', 'ship' and 'truck'.

Requirements

• Python 3
• Numpy 1.14.1
• Keras 2.1.4
• Tensoflow 1.8.0
• Scipy 1.0.0
• Matplotlib 2.1.2

Quick start

You can run an example code of UU learning on benchmark datasets (MNIST, CIFAR-10).

python experiment.py --dataset mnist --mode UU

You can see additional information by adding --help.

Result

After running experiment.py, the test performance figure and log file are made in output/dataset/ by default. The errors are measured by zero-one loss.

Contact: Nan Lu (lu@ms.k.u-tokyo.ac.jp).