0
$\begingroup$

to train a generative adversarial network, a random input is given to the generator to produce a random output, which is given to the discriminatant and, by backpropagation, tweak the generator's parameters.

However, if the generator's parameters are only tweaked with the same random input at each iteration, then the generator will be able to produce only a unique output.

So the question is : is the input constant ? And if not, is there any rule to produce other inputs ?

Thanks

$\endgroup$
0
$\begingroup$

The input to the generator comes from an underlying base distribution of the latent space $p(z)$ we want to map to the data space $p(x)$ . The most common choice is standard normal $z \sim \mathcal{N}(0, 1)$.

The objective here is not to map one single low dimensional vector to the data space (say images) but to learn the true underlying data generating distribution $p(x)$ - say the distribution of all images.

$\endgroup$
2
  • $\begingroup$ So at each iteration a new random input is token with the normal law ? We never take an input (except randomly) multiple times ? $\endgroup$ – Baxlan Dec 13 '19 at 15:53
  • $\begingroup$ Yes. We sample a random Gaussian and send it through. It is almost improbable to sample the same vector twice so there's that. We don't particularly enforce an input being sent multiple times. It just doesn't happen out of randomness. $\endgroup$ – activatedgeek Dec 30 '19 at 18:45

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.