eli5 what Ai generated artwork is and how does someone accomplish this.
In: 0
It really depends on the underlying technology, but I’ll try my best to describe it.
Let’s say I ask you what a sheep is. You know what a sheep is, so I then ask you to draw a picture if a sheep. This is because you’ve seen a sheep before, you know what it looks like, so you can try to replicate it. You can also learn new things; so let’s say I ask you to draw an aardvark, but you’ve never seen an aardvark! So your drawing won’t be that good. But the more times you see an aardvark, the better your drawing will be.
This is a core concept of machine learning. We have a robot, where we show them the word (“sheep”, “aardvark”, etc) and then show them a picture of that thing. After seeing thousands and thousands of pictures, the robot eventually begins to learn how it works.
We can do the same for artwork too; maybe we’ll show it a type of art (watercolor, abstract, geometric, etc), and give it the title of the art, and then show it the actual art piece. Then, after repeatedly seeing thousands of pieces of art, it understands how to make it; we give it the same things to start (let’s say “geometric” as the type and “cardinal” as the title), and it’ll spit something out that it thinks fits best.
This isn’t an accurate explanation, but it should convey the principle of how machine learning works.
Basically, a computer takes a huge amount of samples of art, and generates something new more or less completely at random, either from scratch, or by actually sampling bits of the original dataset.
It’ll then see which if the random products came closest to matching the data, and pick the next seed for it’s random number generator to be close to the most successful one it’s used so far.
It repeats this process millions upon millions of time, each time creating a random number generator weighted slightly closer to matching it’s dataset.
Now you take this principle, and you pair it with some PhD level mathematics in order to optimize the process, pick the right random number generators, evaluate the datasets and stuff like that, and eventually you get something that will coincide close ebough to what you want it to do.
It’s not so much an AI as it is a big pile of data calculus and statistics, that barfs out very carefully filtered and weighted nonsense, such that it comes close enough to what you want it to do, often very effectively.
Latest Answers