Counter-examples generation from a positive unlabeled image dataset

PATTERN RECOGNITION

卷:107

Article

Chiaroni, Florent^1,3  Khodabandelou, Ghazaleh²  Rahal, Mohamed-Cherif¹  Hueber, Nicolas⁴  Dufaux, Frederic³ 

[1] VEDECOM Inst, Dept Delegated Driving VEH08, Percept Team, 23 Bis Allee Marronniers, F-78000 Versailles, France

[2] Univ Paris Est, Lab Images Signaux & Syst Intelligents LISSI, 120 Rue Paul Armangot, F-94400 Vitry Sur Seine, France

[3] Univ Paris Saclay, Lab Signaux & Syst, Cent Supelec, CNRS, 3 Rue Joliot Curie, F-91190 Gif Sur Yvette, France

[4] French German Res Inst St Louis ISL, ELSI Team, 5 Rue Gen Cassagnou, F-68300 St Louis, France

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

DOI  :  10.1016/j.patcog.2020.107527

来源: Elsevier

【 摘 要 】

This paper considers the problem of positive unlabeled (PU) learning. In this context, we propose a two-stage GAN-based model. More specifically, the main contribution is to incorporate a biased PU risk within the standard GAN discriminator loss function. In this manner, the discriminator is constrained to steer the generator to converge towards the unlabeled samples distribution while diverging from the positive samples distribution. Consequently, the proposed model, referred to as D-GAN, exclusively learns the counter-examples distribution without prior knowledge. Experimental results on simple and complex image datasets demonstrate that our approach outperforms state-of-the-art PU methods without prior by overcoming issues such as sensitivity to prior knowledge or first-stage overfitting. (C) 2020 Elsevier Ltd. All rights reserved.

【 授权许可】

Free