Least Squares Generative Adversarial Networks
Генеративные состязательные сети с методом наименьших квадратов
2017-10-01
SCID: 54.1/scwvb7q3
Discuss with AI
CIFAR-10LSGANLSUNLeast Squares Generative Adversarial NetworksPearson X2 divergenceleast squares loss
Figures from the paper
Abstract (AI)
Unsupervised learning with generative adversarial networks (GANs) has proven hugely successful. Regular GANs hypothesize the discriminator as a classifier with the sigmoid cross entropy loss function. However, we found that this loss function may lead to the vanishing gradients problem during the learning process. To overcome such a problem, we propose in this paper the Least Squares Generative Adversarial Networks (LSGANs) which adopt the least squares loss function for the discriminator. We show that minimizing the objective function of LSGAN yields minimizing the Pearson X2 divergence. There are two benefits of LSGANs over regular GANs. First, LSGANs are able to generate higher quality images than regular GANs. Second, LSGANs perform more stable during the learning process. We evaluate LSGANs on LSUN and CIFAR-10 datasets and the experimental results show that the images generated by LSGANs are of better quality than the ones generated by regular GANs. We also conduct two comparison experiments between LSGANs and regular GANs to illustrate the stability of LSGANs.
Key Findings
1
LSGANs demonstrate more stable training behavior than regular GANs, supported by two comparison experiments.
2
LSGANs produce higher-quality generated images than regular GANs on LSUN and CIFAR-10 according to the authors' experiments.
3
Minimizing the LSGAN objective is equivalent to minimizing the Pearson X^2 divergence between model and data distributions.
4
Replacing the sigmoid cross-entropy loss in GAN discriminators with a least squares loss (LSGAN) reduces vanishing gradients during training.
Research Object
Least Squares Generative Adversarial Networks (LSGANs)
Research Subject
Using a least squares loss for the discriminator to improve training stability and generate higher-quality images compared to regular GANs (including relation to minimizing Pearson X^2 divergence)
Publication Details
Publication Date
2017-10-01
Journal
Publisher
ISSN
Cited by
5383
Open access PDF
Access Type
Author Information
Download PDF
Subscribe to digest
References available in scid.ai9
Deep Residual Learning for Image Recognition2016
Very Deep Convolutional Networks for Large-Scale Image Recognition2014
Adam: A Method for Stochastic Optimization2014
Image-to-Image Translation with Conditional Adversarial Networks2017
Learning Multiple Layers of Features from Tiny Images2024
Reducing the Dimensionality of Data with Neural Networks2006
Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network2017
Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks2015
Data Driven VRP: A Neural Network Model to Learn Hidden Preferences for VRP2021
Cited by5
Unpaired Image-to-Image Translation Using Cycle-Consistent Adversarial Networks2017
A survey on Image Data Augmentation for Deep Learning2019
Reconstruction of three-dimensional porous media using generative adversarial neural networks2017
Robust motion in-betweening2020
A Survey on Audio-Video Based Defect Detection Through Deep Learning in Railway Maintenance2022