GAN 이란 무엇인가? 생성 Adversarial Networks 완벽 가이드

The Artist and the Art Critic: How GANs Changed AI Forever

Imagine an ambitious young painter trying to master portraiture. At first, their sketches look nothing like real people—eyes are misaligned, proportions are wildly off, and the colors look bizarre. Now, imagine pairing this painter with a ruthless, hyper-critical art expert whose only job is to look at the artwork and declare, “Real” or “Fake.”

Every time the painter produces a new canvas, the critic inspects it, spots the flaws, and explains what gave it away. The painter takes that feedback, hits the drawing board, and tries harder to fool the critic. Meanwhile, the critic also sharpens their eye, learning to catch even more subtle mistakes. Month after month, year after year, this fierce competition continues. Eventually, the painter becomes so exceptionally skilled that their work is completely indistinguishable from reality.

This high-stakes game is the exact concept behind a GAN, or Generative Adversarial Network. Introduced by computer scientist Ian Goodfellow and his colleagues in 2014, GANs revolutionized artificial intelligence by introducing a way for machines to create entirely new, original data out of thin air.

The Two Halves of a GAN

At its core, a GAN consists of two distinct neural networks locked in an endless, digital duel:

1. The Generator (The Forger): This network takes random noise as its input and tries to transform it into something meaningful, like a realistic photo of a human face, a song, or a 3D object. In the beginning, its outputs are random pixelated noise. But with every training cycle, it adjusts its internal parameters to get better at creating convincing forgeries.
2. The Discriminator (The Critic): This network acts as a detective. It is fed a mix of actual, real-world data (like real photographs) and fake data created by the generator. Its job is to output a probability score determining whether what it is looking at is authentic or synthetic.

The term “adversarial” means opposed or conflicting, which perfectly describes their relationship. They have opposing goals: the generator wants to trick the discriminator, while the discriminator wants to avoid being fooled.

Reaching Equilibrium

The magic of a GAN lies in how they train each other. If the discriminator is too weak, the generator never improves because it gets away with sloppy work. If the discriminator is too strong right out of the gate, it instantly rejects everything the generator makes, starving it of useful feedback.