There's no doubt that GAN is a two-player game, but is it zero-sum?
Based on the definition of a zero-sum game mentioned in the lecture, if L_theta = - L_phi, this game is a zero-sum game. But according to the loss functions of GAN, they don't match the definition of a zero-sum game.
The original one is zero-sum; but in practice, people use the "non-saturating" loss for G, and that one is no longer zero-sum. There are also many more variants of the losses used in practice that we did not mention in the course, and these are usually non-zero-sum.
Is GAN a zero-sum game?
There's no doubt that GAN is a two-player game, but is it zero-sum?
Based on the definition of a zero-sum game mentioned in the lecture, if L_theta = - L_phi, this game is a zero-sum game. But according to the loss functions of GAN, they don't match the definition of a zero-sum game.
1
The original one is zero-sum; but in practice, people use the "non-saturating" loss for G, and that one is no longer zero-sum. There are also many more variants of the losses used in practice that we did not mention in the course, and these are usually non-zero-sum.
and out of curiosity does the loss function we use guarantee we reach nash equilibrium, and is that equilibrium unique?
not at all :) (with exceptions for very simple toy examples--such as the bilinear game min_x max_y xy, but definitely not for DNNs )
Ah ok thanks- why did we discusss nash equilibrium then if we don't actually reach it using the Gan algorithm?
Add comment