It looks similar. In any case, literally none of the literature I've seen on the topic references it. You'd think this would at least be the starting point for most GAN training methods.
-
This Tweet is unavailable.
-
This Tweet is unavailable.
-
There are actually a few GAN training papers which are basically variations of extragradient ascent. See
@AnimaAnandkumar paper which proposes a better alternative: https://arxiv.org/pdf/1905.12103.pdf …
CGD uses mixed Hessian and has nice guarantees that each update is Nash equilibrium of local game. So the two players are in equilibrium in each update. In experiments we found CGD can handle far higher learning rates because of this property. See thread:https://twitter.com/AnimaAnandkumar/status/1205173860284293121?s=20 …
