Fun fact, this loop is kinda how one of the generative ML algorithms works. This algorithm is called Generative Adversarial Networks or GAN.
You have a so-called Generator neural network G that generates something (usually images) from random noise and a Discriminator neural network D that can take images (or whatever you’re generating) as input and outputs whether this is real or fake (not actually in a binary way, but as a continuous value). D is trained on images from G, which should be classified as fake, and real images from a dataset that should be classified as real. G is trained to generate images from random noise vectors that fool D into thinking they’re real. D is, like most neural networks, essentially just a mathematical function so you can just compute how to adjust the generated image to make it appear more real using derivatives.
In the perfect case these 2 networks battle until they reach peak performance. In practice you usually need to do some extra shit to prevent the whole situation from crashing and burning. What often happens, for instance, is that D becomes so good that it doesn’t provide any useful feedback anymore. It sees the generated images as 100% fake, meaning there’s no longer an obvious way to alter the generated image to make it seem more real.
Fun fact, this loop is kinda how one of the generative ML algorithms works. This algorithm is called Generative Adversarial Networks or GAN.
You have a so-called Generator neural network G that generates something (usually images) from random noise and a Discriminator neural network D that can take images (or whatever you’re generating) as input and outputs whether this is real or fake (not actually in a binary way, but as a continuous value). D is trained on images from G, which should be classified as fake, and real images from a dataset that should be classified as real. G is trained to generate images from random noise vectors that fool D into thinking they’re real. D is, like most neural networks, essentially just a mathematical function so you can just compute how to adjust the generated image to make it appear more real using derivatives.
In the perfect case these 2 networks battle until they reach peak performance. In practice you usually need to do some extra shit to prevent the whole situation from crashing and burning. What often happens, for instance, is that D becomes so good that it doesn’t provide any useful feedback anymore. It sees the generated images as 100% fake, meaning there’s no longer an obvious way to alter the generated image to make it seem more real.
Sorry for the infodump :3