Deepfakes are an AI technology that makes it look like someone said or did something they never actually said or did. You’ve probably seen a few online. These videos may look real, but they are completely fake. Behind this technology is a powerful deep learning model called a Generative Adversarial Network, or GAN for short.

To me, GANs work like a fascinating game, especially the way they’re trained. Come with me, let me explain.

A GAN is made up of two networks:

• 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗼𝗿 creates fake (synthetic) data samples
• 𝗗𝗶𝘀𝗰𝗿𝗶𝗺𝗶𝗻𝗮𝘁𝗼𝗿 tries to figure out if a sample is real or fake

They act like two players in a cat and mouse game, each one trying to outsmart the other.

At first, the Discriminator is trained on real data (the training data). Once it has a good understanding of what real data looks like, the Generator is brought in.

Now here’s the fun part. The Generator is not trained with the real data. Instead, it is given some random noise (just numbers) as input. Its challenge is to turn that random noise into something so realistic that it fools the Discriminator into thinking it is real data.

𝗦𝗼, 𝗵𝗼𝘄 𝗱𝗼𝗲𝘀 𝘁𝗵𝗲 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗼𝗿 𝗲𝘃𝗲𝗻𝘁𝘂𝗮𝗹𝗹𝘆 𝗽𝗿𝗼𝗱𝘂𝗰𝗲 𝗱𝗮𝘁𝗮 𝘁𝗵𝗮𝘁 𝗹𝗼𝗼𝗸𝘀 𝗿𝗲𝗮𝗹 𝗲𝗻𝗼𝘂𝗴𝗵?

That’s the beauty of the setup. Every time the Discriminator catches a fake, the Generator learns from that mistake and tries again. The Discriminator gets better at spotting fakes, and the Generator gets better at creating more convincing ones. They both improve by competing with each other.

---------------

Pre-order my book "AI, Machine Learning, Deep Learning: From Novice to Pro" here: https://a.co/d/4cRP3oM