ML//GAN//mode collapse

Mode collapse is a training failure of GANs in which the generator finds one output that fools the discriminator and keeps producing only that, and it is the instability anyone training a GAN has to watch for. All generated images then look the same: the model has collapsed onto a single mode of the distribution.

Mode collapse is a training failure of GANs in which the generator finds one output that fools the discriminator and keeps producing only that, and it is the instability anyone training a GAN has to watch for. All generated images then look the same: the model has collapsed onto a single mode of the distribution.

It is the fundamental training instability of GANs. Many hacks mitigate it, and none has fully solved it.