1. Generative Models

1-1. Generative models vs Discriminative models

• Discriminative models

  • $X→Y$
  • Feature → Class
  • $P(Y|X)$

• Generative models

  • $\xi, Y → X$
  • noise, Class → Features
  • $P(X|Y)$

1-2. Generative Models

• Variational Autoencoders (VAE)
  • Encoder → latent space → Decoder
• Generative Adversarial Networks
  • Generator → Discriminator

The Generator’s role is in some sense similar to the decoder in the VAE. The difference is that there’s no guiding encoder that determines what the noise vector should look like that’s input into the generator.

2. Real Life GANs

2-1. Application

• Image Translation

2-2. Major Companies using GANs

• Adobe → next gen photoshop
• Google → Text generation
• IBM → Data augmentation
• snapchat, ticktock - > image filters
• Disney → super resolution

3. Intuition Behind GANs