Is Generator Conditioning Causally Related to GAN Performance?

Is Generator Conditioning Causally Related to GAN Performance?

abs：

Recent work (Pennington et al., 2017) suggests

that controlling the entire distribution of Jacobian

singular values is an important design considera-

tion in deep learning. Motivated by this, we study

the distribution of singular values of the Jacobian

of the generator in Generative Adversarial Net-

works (GANs). We find that this Jacobian gen-

erally becomes ill-conditioned at the beginning

of training. Moreover, we find that the average

(with z ∼ p(z)) conditioning of the generator

is highly predictive of two other ad-hoc metrics

for measuring the “quality” of trained GANs: the

Inception Score and the Frechet Inception Dis-

tance (FID). We test the hypothesis that this re-

lationship is causal by proposing a “regulariza-

tion” technique (called Jacobian Clamping) that

softly penalizes the condition number of the gen-

erator Jacobian. Jacobian Clamping improves

the mean Inception Score and the mean FID for

GANs trained on several datasets. It also greatly

reduces inter-run variance of the aforementioned

scores, addressing (at least partially) one of the

main criticisms of GANs

https://www.arxiv-vanity.com/papers/1802.08768/