Digest and Test: TequilaGAN—Identifying GAN Samples

In this chapter, we investigated the numerical properties of samples produced with adversarial methods, especially GANs. We showed that fake samples have properties that are barely noticed within the visuals of samples, namely the fact that due to stochastic gradient descent and the requirements of differentiability, fake samples smoothly approximate the dominating modes of the distribution. We analyzed statistical measures of divergence between real data and other data, and the results showed that even in simple cases—for instance, distribution of pixel intensities—the divergence between training data and fake data is large with respect to test data.

Although not common practice, we could possibly circumvent the difference in support between the real and fake data by training generators that explicitly sample a distribution that replicates the support of the real data, for instance, 256 values in the case of discretized images. These are topics that are not limited to GANs and remain to be explored in the larger domain of verified artificial intelligence.

Test your understanding

Put your knowledge to the test with a quiz designed to reinforce your understanding.