What discriminator? And what is the "real" in this case?to_the_sun wrote:Such hostility. I fear this post is no longer on a productive track. Oh well…
I do play several hours a day. And even when the AIs are making better music than I ever could, I still will be, because I find it enjoyable not because it might make me money.
True that new sounds are not necessarily usable. That's the point of the adversarial network. It refines the results until the discriminator can't tell the difference between real and generated.
So, what you will end up is variations of something that already existed. There are synths that already have this incorporated. Not exactly what you may looking for, but there are. And it's no big deal.to_the_sun wrote: No, machine learning doesn't make choices. That's not how it works. But what's "good" is determined by the training examples
Believe in just half of what you see, and even that half, you should keep questioning. The problem is always "the choices". Humans learn basically by trial and error, But machines can't sepoarate what is "right" from what is "wrong" (because that's subjective, right from start). So, they can be guided by someone, but that will just lead to results that are considered good buy that "someone" - not necessarily by others.to_the_sun wrote: Oh, and yes, the AIs are beginning to program themselves: https://www.youtube.com/watch?v=hVv68aHYSs4
Remember that we are not dealing with some objective and measurable reality.