ML - Generative Adversarial Network (GAN)

Generative Adversarial Network

ANN that uses Adversarial Learning with generator and discriminator to make high qulity images.

 

Adversarial Learning Process

Generator is trained to produces Real-like images.

Discriminator is trained to distinguish fake images well.

Train Discriminator first and Generator next with backpropagation.

 

Object of GAN Training

D(x) maximize it's reward.

G(z) minimize D(x)'s reward.

In implementation, GAN uses BCE as cost function.

 

Changing Object Function

Limitation of Vanilla GAN

Non-Convergence Problem

When both sides develop little by little, it becomes Nash equilibrium, but when one side becomes stronger, the other side cannot develop.

Mode-Collapse Problem

If it is not properly trained, it does not create various images, but it continues to create similar images.

 

Variants of GAN

Deep Convolutional GAN (DCGAN)

Replace FC hidden layers (Generator) with Convolutions

 

Semi-Supervised GAN (SGAN)

The generated image is controlled using a one-hot vector having information on the generated image.

 

Other GANs