GAN (Generative Adversarial Networks)
Variants
InfoGAN [14]
cGANs [15] (Conditional Gan)
CycleGAN [16]
f-GAN [17]
WGAN [18]
WGAN-GP [19]
LS-GAN [20]
Generative algorithms
Explicit density model
Implicit density model
GANs vs other generative
algorithms
GANs can parallelize the generation, which is impossible for other generative algorithms such as PixelCNN [104] and fully visible belief networks
(FVBNs) [105], [106]
The generator design has few restrictions.
GANs are subjectively thought to produce better examples than other methods
Structure
Original GANs
Laplacian generative adversarial networks (LAPGAN) and SinGAN
Deep convolutional generative adversarial networks (DCGANs)
Progressive GAN
Self-Attention Generative Adversarial Network (SAGAN)
BigGANs and StyleGAN
Hybrids of autoencoders and GANs
Multi-discriminator learning
Multi-generator learning
Multi-GAN learning
Evaluation metrics for GANs
Inception Score (IS)
Mode score (MS)
Frechet Inception Distance (FID)
Multi-scale structural similarity (MS-SSIM)
Task driven by GANs
Semi-Supervised Learning
Transfer learning
Reinforcement learning
Multi-modal learning
active learning, online learning, ensemble learning, zero-shot learning, and multi-task learning
APPLICATIONS
Image processing and computer vision
Super-resolution (SR)
Image synthesis and manipulation
Texture synthesis
Object detection
Video applications
Sequential data
Natural language processing (NLP)
Music
Speech and Audio
Others
Medical field
Data science
OPEN RESEARCH PROBLEMS
GANs for discrete data
New Divergences
Estimation uncertainty
Theory
Generative Recurrent Adversarial Networks (GRAN)