Perceptron / Neuron

Neural Network

Goal: optimise some loss function (mean square error, cross-entropy...)

Generative Adversarial Network (GAN)

We want to:

  • Learn the distribution that generates the data p(x).
  • Generate new samples from p(x)

How we do it:

  • Two CNNs: Generator G and Discriminator D.
  • We train G to generate new "fake" samples to trick D.
  • We train D to distinguish fake and real samples.
  • Training ends when there is a "stalemate".

Generative Adversarial Network (GAN)

Example

GAN dissection

By arnauqb

GAN dissection

  • 609