Please enable JavaScript.
Coggle requires JavaScript to display documents.
DL(24) - Generative Adversarial Network - Coggle Diagram
DL(24) - Generative Adversarial Network
Game-theoretic approach
sample from a simple distribution, e.g random noise, and learn transformation (generator network) to training distribution
Trasformation should be complex enough: a deep neural network as generator network!
learn to generate from training distribution through 2-player game
Structure
Discriminator Network
: try to distinguish between real images (target = 1) and fake images (target = 0)
The
training
procedure is:
gradient
ascent
on Discriminator
gradient
descent
on Generator
Generator Network
: try to fool the
Discriminator
by generating real-looking images
VAEs VS GANs
Generative Adversarial Networks
+++ Best samples
--- Can be tricky and unstable to train, no inference queries
Game-theoretic approach
Variational Autoencoders
+++ Useful latent representation that allows for inference queries
--- Usually sample quality is not the best
Optimize variational lower bound on likelihood