What is a Generative Adversarial Network (GAN)? How does it generate new data?

Okay, let's talk about Generative Adversarial Networks (GANs).

Forget the complex jargon. You can imagine a GAN as a story of intense rivalry and mutual growth between a "master forger" and an "appraisal expert."

These two roles are actually two independent neural networks:

1. Generator: This is the master forger. His goal is to create highly realistic fakes from scratch. For example, if you show him ten thousand cat photos, his task is to draw a brand new cat that looks incredibly real. Initially, his drawings might be an unrecognizable mess.

2. Discriminator: This is the appraisal expert. His task is very straightforward: to determine if an item is "genuine" or a "fake." In this example, it means judging whether an image is a real cat (from your dataset) or a fake cat drawn by the master forger.

How do they "fight" and learn?

This process is like a continuous battle:

• Round One:

  • The master forger (Generator) draws a very ugly fake cat and mixes it in with a bunch of real cat photos.
  • The appraisal expert (Discriminator) looks at them, and because the fake cat is so poorly drawn, he easily points it out: "Ha, this one's a fake!"

• Learning and Evolution:

  • The master forger receives feedback: He learns that his cat drawing was exposed because of "certain features" (e.g., the ears were too square, the fur wasn't smooth). So, he learns his lesson and pays special attention to improving that aspect next time he draws.
  • The appraisal expert also receives feedback: He is "rewarded" for successfully identifying the fake, which boosts his confidence and clarifies what a real cat should look like.

• Round Two:

  • The master forger draws a fake cat that's a bit better than last time.
  • The appraisal expert might need to take a closer look this time, but he still identifies it: "Hmm, improved, but still a fake."

• Continuous Cycle...

  • This process repeats thousands of times. The master forger's drawing skills become increasingly sophisticated, evolving from "unrecognizable" to "somewhat similar" and then to "very realistic."
  • Meanwhile, the appraisal expert's eye for detail also becomes sharper; he can notice increasingly subtle flaws.

What is the final outcome?

The ultimate goal of this "arms race" is: the fake cats drawn by the master forger become so good that the appraisal expert can no longer distinguish them from real ones. When the appraisal expert sees a fake cat image, his probability of guessing correctly or incorrectly is only 50%, just like a random guess.

At this point, we consider the master forger (Generator) to have "mastered his craft."

How does it generate "new" data?

Once our "master forger" (Generator) is well-trained, he becomes a true artist.

You no longer need to show him real cats. You just need to give him some random "inspiration" (technically called "random noise," which you can imagine as a jumble of disorganized pixels), and he can use the "essence of cat generation" he's learned to transform this random inspiration into a brand new, high-quality, one-of-a-kind cat image that has never existed before.

In summary:

A GAN operates through a model where a "forger" and an "appraiser" compete and improve together. This adversarial training ultimately compels the "forger" to master the core ability of creating realistic data, thereby enabling the generation of entirely new content that was not present in the original dataset.

So, the next time you see non-existent human faces or AI-generated artworks, it's very likely that a pair of "master forger" and "appraisal expert" are behind them.