Generative Adversarial Networks (GANs)

What is GAN?

= A framework to train generative models via adversarial process.

Architecture

• Generator (G)
  • Creates fake data from random noise
  • Takes noise vector z (sampled from e.g., Gaussian) → outputs synthetic data
• Discriminator (D)
  • Tries to distinguish real data from fake data
  • Takes data (real or fake) → outputs probability of being real

How it works

• "adversarial"

  • G tries to fool D by producing realistic samples
  • D tries to correctly classify real vs fake
  • Both networks train simultaneously in a minimax game

Training Process

1. Sample real data batch
2. Sample noise batch
3. Train D on real and fake data
4. Train G to fool D
5. Repeat until convergence

Loss Functions

• Discriminator loss:

• Generator loss:

Challenges

• Training instability
• Mode collapse (generator produces limited diversity)
• Balancing G and D power