Counter-examples generation from a positive unlabeled image dataset

作者：

Highlights：

• It is possible to constrain a Generative Adversarial Network (GAN) in order to capture the counter-examples distribution from a Positive Unlabeled (PU) dataset without prior.

• Some deep learning regularization techniques can manage several minibatch sample distributions.

• The proposed approach incorporates a biased PU training loss function in the original GAN discriminator loss function.

摘要

•It is possible to constrain a Generative Adversarial Network (GAN) in order to capture the counter-examples distribution from a Positive Unlabeled (PU) dataset without prior.•Some deep learning regularization techniques can manage several minibatch sample distributions.•The proposed approach incorporates a biased PU training loss function in the original GAN discriminator loss function.

论文关键词：Generative adversarial networks (GANs),Generative models,Semi-supervised learning,Partially supervised learning,Deep learning

论文评审过程：Received 16 October 2019, Revised 22 May 2020, Accepted 30 June 2020, Available online 1 July 2020, Version of Record 6 July 2020.

论文官网地址：https://doi.org/10.1016/j.patcog.2020.107527