Генератор и дискриминатор: игра

A GAN is two neural networks playing a game: one invents fake data, the other tries to catch it. Their rivalry is what makes both improve. 🎭

The generator starts with random noise and shapes it into something that looks like real data, such as a fake face. It never sees the real images directly.

The discriminator is a judge. It looks at a sample and outputs one number: how likely it thinks that sample is real rather than fake.

You feed the generator a vector of random numbers called the latent vector. Different vectors produce different outputs, so noise becomes the seed of creativity.

Think of a forger making fake bills and a detective spotting them. Each one pushes the other to get sharper. That tug-of-war is the heart of a GAN.

The discriminator trains on labels: real images get 1 and generated images get 0. It learns to tell the two apart with confidence.

The generator wins when the discriminator labels its fakes as real. So its goal is to fool the judge, not to copy any single image exactly.

Their objectives are opposite, which is why we call it adversarial. The discriminator wants to be right; the generator wants it to be wrong.

As the discriminator gets pickier, the generator must make better fakes to keep fooling it. This feedback loop lifts both networks over time.

Because the players have separate goals, each network gets its own optimizer and updates on its own turn during training.

Training ideally ends when the discriminator can only guess, near 50 percent. At that point the fakes are practically indistinguishable from real data.

Let's make sure the roles are clear.

You met both players: a generator that forges data from noise and a discriminator that judges it. Their rivalry drives both to improve. 🎉