February 2019
Intermediate to advanced
386 pages
9h 54m
English
Let's suppose that we have a GAN that has been properly trained by using a dataset X, drawn from pdata(x). Goodfellow, et al., proved that, given the generator distribution pg(x), the optimal discriminator is as follows:

The global objective can be rewritten by using the optimal discriminator:

We can now expand the previous expression:

Now, let's consider the Kullback-Leibler divergence between two distributions, a and b:
Read now
Unlock full access