GAN - Simplified

GAN stands for Generative Adversial Network.

GAN consists of 2 neural networks :

1.Generator

2.Discriminator

GENERATOR

Generator is a neural network trained to generate images.

Using random input, it starts generating fake images at first, then slowly learns to generate fake images that look almost same as our input images.

DISCRIMINATOR

Discriminator is a neural network trained to differentiate Real & Fake images.

It is trained on 2 types of data:

1.Real data from input

2.Fake data from generator

TRAINING GOALS

Goal of training the generator is to create images that resemble input

Goal of training the discriminator is to differentiate real & generated data

Goal of training the GAN is to make the generator generate data that the discriminator classifies as real.

TRAINING PROCESS

Generator & Discriminator are trained alternatively

The following process is repeated iteratively :

1.Generator generates fake data.

2.Discriminator classifies it as fake.

3.Feedback is given to Generator.

4.Generator generates better fake data

TRAINING DURATION

As the generator becomes better at creating fake images that look real, discriminator becomes worse at differentiating real & fake images and wrongly classifies the fake generated data as real. This is the stage where we stop training.

Now we have a GAN ready to generate new data !