Learning with A Generative Adversarial Network From a Positive Unlabeled Dataset for Image Classification

In this paper, we propose a new approach which addresses the Positive Unlabeled learning challenge for image classification. Its functioning is based on GAN abilities in order to generate fake images samples whose distribution gets closer to negative samples distribution included in the unlabeled dataset available, while being different to the distribution of the unlabeled positive samples. Then we train a CNN classifier with the positive samples and the fake generated samples, as it would be done with a classic Positive Negative dataset. The tests performed on three different image classification datasets show that the system is stable up to an acceptable fraction of positive samples present in the unlabeled dataset. Although very different, this method outperforms the state of the art PU learning on the RGB dataset CIFAR-10.