Summary
•
propose a regularization term to stabilize GANs by adjusting the eigenvalues of the Jacobian matrix.
•
Strength
Weakness
Improvement Points
•
use alternating instead of simultaneous gradient descent
•
don’t use momentum (θ_t+1 = β˜θ_t + (1 − β)∇θ_t, θt+1 = θt + η˜θt+1, β=0)
•
use regularization to stabilize the training
•
simple zero-centered gradient penalties for the discriminator yield excellent results
•
progressively growing architectures might be not all that important when using a good regularizer
