[LOCKSUIT]

GANs are said to have 2 networks compete against each other, improving each other, one gets better at generating ex. realistic images, the other gets better at identifying fake images. But I think I found proof it is just a normal network in the end, not 2 and not anything new, just a new way of doing the same thing, like destroying a building using a missile versus using lava...

> How can either suddenly "get better"? We tell the cop net it got fooled and we know it is generated, not real image. Or tell the scammer it got caught, didn't fool the other net. But telling either of them this is saying that the image could be better, mr. thief, or the cop's understanding of what is better could be better, mr. cop! Same thing! And, telling them this doesn't tell them HOW to make the image better! You NEED to have references! So in the end a single network gets better by trying to recreate the original images using the least needed training i.e. you are good at predicting reality / future questions never seen before using just a little bit of the past.

Prove me wrong.

Reading this one link here and it's nonsense as usual.
https://towardsdatascience.com/how-gans-really-work-2e1db1f407bb
One part though says:
"We want to generate new samples, new situations which describe the same high-level features as X. And that%u2019s exactly what networks does : we use their weakness for our profit."
hmm.....

[LOCKSUIT]

i found below it says in an article:
"In order to train the generator, the discriminator have to tell the generator how to tweak that generated image to be more realistic.

The generator must ask for suggestions from the discriminator!

Intuitively, the discriminator tells how much to tweak each pixel in order to make the image a little bit more realistic.

Technically, you do that by back-propagating the gradients of the discriminator’s output with respect to the generated image. That way, you will have gradients tensor that have the same shape as the image.

If you add these gradients to the generated image, it will make the image more realistic (in the discriminator’s eyes).

But we are not just going to add these gradients to the image.

Instead, we will back-propagate these image gradients further into all the weights that made up the generator so that the generator learns how to generate this new image."

[LOCKSUIT]

not sure what you mean but I propose another way to do what GANs do, without a 2nd net. What you need do is take your existing net and look at which tree branches (think trie, markov chain) don't have enough statistics, and tell it it needs to tweak that branch or collect more data. For example 2 letter branch in a trie for text would be observing aa, ab, af, aa, zu, aa, aa, aq, aa.....and here we see aa is common, and we may not need to see any more 2 letter features because we see the trend now, but 66 letters we don't yet ex. hdfherdf.......zzzhdqwwoi.....no reoccurences yet here....

cus i mean, all a GAN can do is either say tweak the net or where to tweak it, which is equivalent to saying collect more data or collect specific data......otherwise a GAN would need to do some other function like Backoff or Semantics i.e. introduce other means to improve the net, not just tell it where it's generated image is lacking. See my book from last month version to get what i mean: