Generative Adversarial Network (GAN)s

Learning from the google developer page at this link .

Structure of GAN

GAN consists of two parts:

Generator:
- is a neural network.
- Training is done to produce output, that is as much plausible.
- generator output is connected to the discriminator input
Discriminator:
- is a neural network.
- Penalizes the Generator, if the output from generator is not plausible.
- through backpropagation, the discrimination provides signals that generator uses to update its weights.

has two sources of data: real data and fake data.
Real data are the positive examples from the real world.
Fake data are the outputs generated from the generative part of GAN.
During discriminator training, the weights of generator is kept constant, i.e. generator does not train.
During Training:
- Discriminator classifies real and fake data that is input into it
- it is penalized for misclassifying the inputs
- it computes the loss, and using backpropagation updates its weight

creates data and feeds to the discriminator
will reach success when discriminator classifies all outputs as true/real
The generator loss penalizes the generator for producing a sample that the discriminator network classifies as fake.
During training of the generator, the backpropagation starts at the output and flows back through the discriminator to the generator
But only the weights of the generator are changed. The weights of the discriminator is unchanged.

For GAN training, two distinct training must be done. And there are some tricks to updating weight as mentioned above.
Following is the summary of the overall steps:
- Train discriminator for one or more epochs
- Train generator for one or more epochs
- Repeat
As the generator gets better with training, discriminator becomes worse as it cannot tell difference between the data generated or real data. If the generator is successful, the discriminator has 50% accuracy
There are two major algorithms for loss function computation. Refer to page this for more detail.s