Not controlled or controlled by GAN?

Question

From some sources I hear that Generative adversarial networks is an uncontrolled ML, but I don't understand it. Are generative adversarial networks virtually uncontrollable?

1) 2-Class Real-vs-Fake Case

Indeed, you need to provide training data for the discriminator, and this should be "real" data, that is, data that I would call fe 1. Although no one marks the data explicitly, this is done implicitly, introducing the discriminator in the first steps with the training data that you say that the discriminator is genuine. Thus, you somehow tell the discriminator about the marking of training data. Conversely, marking the noise data generated in the first steps of the generator, which the generator knows that they are unreliable.

2) Multi-class case

But this is very strange in the case of several classes. Descriptions must be provided in the training data. The obvious contradiction is that it provides an answer to the uncontrolled ML algorithm.

Answer 1

GANs are unsupervised learning algorithms that use controlled loss as part of learning. It turns out later that you hanged yourself.

When we talk about supervised learning, we usually talk about learning to predict the label associated with the data. The goal is for the model to summarize new data.

In the case of GAN, you do not have any of these components. Data comes without labels, and we are not trying to generalize any forecasts to new data. The goal is for the GAN to simulate what the data looks like (such as a density estimate) and be able to generate new examples of what it learned.

GAN , , / , . , . GAN, .

. " " ( ), SVM , (.. ), .