# Art With Minimal Human Input? An Image Classifier Judges A Genetic Algorithm. | Artificial intelligence

For anyone deeply allergic to Medium or has used their free articles for the month and can’t circumvent the paywall (cough, incognito), here are the images from the article and a section describing the bot’s structure:

AI generated images

So what’s going on here? The shorthand description is that these images are created by a genetic algorithm controlled painting process with selection fitness provided by an image classifier neural network. That’s lot of concentrated jargon to digest so I’ll step through it and try to translate it into non-technobabble English.

The individual images in the group are created by drawing simple geometric shapes: ellipses, capsules, triangles, rectangles. Each drawing is described by a series of numbers that give the order, position, colour and so on of these shapes.

A genetic algorithm is a process that can take a set of parameters (like the numbers describing our drawings) and apply the evolutionary concepts of selection and reproduction with variation to search for combinations that produce the best “fitness” by whatever selection criteria we choose. In terms of our drawings this reproduction and variation is a process of copying the numbers from one of the better scoring images and then randomly changing a few values or adding new random shapes. Roughly speaking, we can use the genetic algorithm to “breed” images that display characteristics we want.

This concept of “what we want” is provided by the image classifier network. classifiers are increasingly common in apps and places like Google’s image search and you probably use them without even knowing. They read images and attempt to identify the subject matter. Behind the scenes most of these classifiers will produce a set of confidence scores for a number of possible subjects. It’s these confidence scores that this art generator uses as fitness for the purposes of selecting images to “breed”.

To sum up, I ask for pictures on the theme of “hotdog” and the genetic algorithm will generate populations of images and use the classifier’s confidence that there’s a hotdog in them to guide an evolutionary search for a more “hotdoggy” image.