Generative AI & the senses
A GAN (generative adversarial network) is a pair of neural networks that train by competing: one creates fake examples and the other tries to spot the fakes.
Invented in 2014 by Ian Goodfellow and colleagues, a GAN sets up a contest. The generator produces images from random numbers. The discriminator looks at a mix of real and generated images and guesses which are real. Each network's mistakes teach the other. Over many rounds, the generator's fakes get good enough to fool the discriminator.
The word 'adversarial' captures the rivalry. Neither network is told what a good image looks like; the pressure to outsmart the other does all the teaching. This approach produced the first strikingly realistic AI faces and powered the early deepfake tools.
GANs are tricky to train: the contest can collapse, with the generator producing the same few images. By the early 2020s, diffusion models mostly replaced them for image generation. GANs are still used for tasks like sharpening images and creating synthetic training data.
A forger keeps painting fake banknotes while a detective keeps learning to spot them; the forger ends up producing notes nobody can tell from real ones.