Some time ago an AI which could convert an image of someone into a drawing would be an idea of science fiction, but in the last year they suddenly blew up on the internet and became widespread, encompassing so many different variations of the images it can generate.
Why did this type of artificial intelligence develop so suddenly? Did someone code an all-purpose source program that enabled this or something?
In: 8
Something not already mentioned is that while the images being shared have suddenly begun appearing, the AI behind them didn’t come out of nowhere. People have been working on the pieces for years.
For example, one of the big reasons these image generation systems are possible is the previous work on image classification. A huge number of training images along with an incredible manual effort enabled some of these neural networks to be able to tell if an image contains, say, a hot dog. It started out simple, but now the classification systems can give much more accurate responses, including facial recognition.
So the next logical step was to use these classification systems to train image generation algorithms using what is called a Generative Adversarial Network (GAN). Adversarial is the key point. A rough simplification is a network is seeded with some kind of random noise, then random changes are made to the image and tested by the classifier. Images that score higher for the target classification are fed back into the system in a kind of positive feedback loop. Meanwhile, the classification system gets better at spotting fake images. This continues until the generator produces images that are highly successful at fooling the classifier.
Latest Answers