Why add regularization term to W-Loss function?

Course Q&A Generative Adversarial Networks (GANS) Build Basic Generative Adversarial Networks
1

image
image1059×115 27.6 KB

Here why we’re adding the regularization term to the W-Loss function?
I understand that Critic wants to maximize W-Loss and we enforce L-1 continuity by adding regularization term, which gets bigger if norm of critic is bigger than 1.

But Isn’t that if norm of gradient of critic is bigger than 1, then it actually increasing loss function, thus benefiting Critic? Isn’t that if we want penalize lshouldn’t we subtract it?

2

The additional added regularization helps in keeping the norm of the gradient of the critic close to 1. This is the same concept as normal regularization where addition of a term is performed to clip the weights, the determining factor is lambda hyper-parameter!