@article{Hao,
author = {Hao, J and Wang, C and Zhang, H and Yang, G},
title = {Annealing Genetic GAN for Minority Oversampling},
url = {http://arxiv.org/abs/2008.01967v1},
}

Download

TY  - JOUR
AB  - The key to overcome class imbalance problems is to capture the distributionof minority class accurately. Generative Adversarial Networks (GANs) have shownsome potentials to tackle class imbalance problems due to their capability ofreproducing data distributions given ample training data samples. However, thescarce samples of one or more classes still pose a great challenge for GANs tolearn accurate distributions for the minority classes. In this work, we proposean Annealing Genetic GAN (AGGAN) method, which aims to reproduce thedistributions closest to the ones of the minority classes using only limiteddata samples. Our AGGAN renovates the training of GANs as an evolutionaryprocess that incorporates the mechanism of simulated annealing. In particular,the generator uses different training strategies to generate multiple offspringand retain the best. Then, we use the Metropolis criterion in the simulatedannealing to decide whether we should update the best offspring for thegenerator. As the Metropolis criterion allows a certain chance to accept theworse solutions, it enables our AGGAN steering away from the local optimum.According to both theoretical analysis and experimental studies on multipleimbalanced image datasets, we prove that the proposed training strategy canenable our AGGAN to reproduce the distributions of minority classes from scarcesamples and provide an effective and robust solution for the class imbalanceproblem.
AU  - Hao,J
AU  - Wang,C
AU  - Zhang,H
AU  - Yang,G
TI  - Annealing Genetic GAN for Minority Oversampling
UR  - http://arxiv.org/abs/2008.01967v1
ER  - 

Download