I heard using hinge loss is a way to prevent mode collapse. Hinge loss seems like stop learning class which is fitted enough. It is not clear fact though, hinge loss acts like early stopping strategy. Evade model remembering data and induce to fit unfamilliar content.
Some GAN use this another loss function, low gradient at high score. Same idea of hinge loss, this loss function also good at mode collapse. But, GAN is still suffered from the problem. It isn't sufficient to improve?
I heard using hinge loss is a way to prevent mode collapse. Hinge loss seems like stop learning class which is fitted enough. It is not clear fact though, hinge loss acts like early stopping strategy. Evade model remembering data and induce to fit unfamilliar content.
Some GAN use this another loss function, low gradient at high score. Same idea of hinge loss, this loss function also good at mode collapse. But, GAN is still suffered from the problem. It isn't sufficient to improve?