Since I don't know where to ask, so I put the question here.
After reading the paper and compare with the Gan version from Goodfellow. I realize that there is a major difference between MUSE and GAN (let's call the version from Goodfellow as GAN).
In GAN, first, one will sample a real image from the real vector spaces, then sample the corresponding fake image from the fake vector space. Actually, the fake image is a noisy version of the corresponding real image, we know exactly what to sample from the fake vector space. Then the goal of the whole training process is to try to make the generator to generate the fake image as similar to the real image as possible, i.e the embedding of this fake image is close to the real image.
But it's not the case of MUSE, right? in MUSE, when sampling an instance from the source space, we don't know which one is the "noisy" version of that instance in the target space. That's why the only way to do this is to randomly sample from the target space.
So to conclude, my question (or you can call a clarification) is that the major difference between MUSE and GAN is that GAN will selectively sample from the fake space given the real image, and MUSE will randomly sample from the target space, is it correct or not?
Since I don't know where to ask, so I put the question here.
After reading the paper and compare with the Gan version from Goodfellow. I realize that there is a major difference between MUSE and GAN (let's call the version from Goodfellow as GAN).
In GAN, first, one will sample a real image from the real vector spaces, then sample the corresponding fake image from the fake vector space. Actually, the fake image is a noisy version of the corresponding real image, we know exactly what to sample from the fake vector space. Then the goal of the whole training process is to try to make the generator to generate the fake image as similar to the real image as possible, i.e the embedding of this fake image is close to the real image.
But it's not the case of MUSE, right? in MUSE, when sampling an instance from the source space, we don't know which one is the "noisy" version of that instance in the target space. That's why the only way to do this is to randomly sample from the target space.
So to conclude, my question (or you can call a clarification) is that the major difference between MUSE and GAN is that GAN will selectively sample from the fake space given the real image, and MUSE will randomly sample from the target space, is it correct or not?
Thanks, :)