GAN

Fully Convolutional

• Remove Max Pooling
  • Use strided UpConv
• Remove FC layers
  • Hurts convergence in non-classification
• Normalisation tricks
  • Batch normalisation
    • Batches of 0 mean and variance 1
  • Leaky ReLu

Stages

Generator, G

• Synthesise ‘fake’ images
• From noise

Discriminator, D

• Discriminator is a classifier #classification
  • Is image fake or real

]

Training

Code Vector Math for Control

• Do AM to derive code for an image