Is Generator Conditioning Causally Related to GAN Performance?

Is Generator Conditioning Causally Related to GAN Performance?

abs：

Recent work (Pennington et al., 2017) suggeststhat controlling the entire distribution of Jacobiansingular values is an important design considera-tion in deep learning. Motivated by this, we studythe distribution of singular values of the Jacobianof the generator in Generative Adversarial Net-works (GANs). We find that this Jacobian gen-erally becomes ill-conditioned at the beginningof training. Moreover, we find that the average(with z ∼ p(z)) conditioning of the generatoris highly predictive of two other ad-hoc metricsfor measuring the “quality” of trained GANs: theInception 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) thatsoftly penalizes the condition number of the gen-erator Jacobian. Jacobian Clamping improvesthe mean Inception Score and the mean FID forGANs trained on several datasets. It also greatlyreduces inter-run variance of the aforementionedscores, addressing (at least partially) one of themain

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