Like who is and where is this stuff generated? Does the ai smash together an image search? Do they direct the references? eli5.
In: 1
First, the neural network is trained on a lot of example data, then the neutral network takes in seed data of some kind and generates an image from that.
The specifics can very greatly depending on type of neutral network. One kind that’s pretty popular these is Generative Adversarial Networks or a “GAN” these work by using 2 neural networks, 1 which generates fake images, and a second one which tries to determine if the images generated are real or not. So, you have one robot trying to fool the other robot that it’s not a robot.
Think of AI learning as like… a photo copy machine.
The copy machine will scan an image and to its best to print out a copy. But the copy isn’t exactly perfect. There’s going to be a few imperfections. Blurs, smudges, blots, and such.
Let’s say I wanted to have it print out the Mona Lisa, but I can’t actually put the Mona Lisa into the machine. I need to ‘teach’ my copy machine how to make the Mona Lisa.
I start by simply pressing copy twice on the machine. I’m going to get two prints that are mostly blank paper. But they aren’t exactly the same. One print will have different imperfections than the other.
I can look at those two prints and compare them to the Mona Lisa. Which print has imperfections, blots, smudges, and such, which kind of looks more like the Mona Lisa than the other.
Obviously, neither print is going to look like the Mona Lisa, but I can look over the prints very carefully and find that one print is closer to my goal than the other.
I can throw away the worst print and keep the better one. I’ll then make a copy of the better print.
I once again have two prints that aren’t exactly the same. I already know what one of those prints looks like compared to the Mona Lisa. I need to check the other print and see if it’s any better.
Maybe the new print is better. Maybe it’s worse. It’s already had a head start, as I’ve given it the best starting image I’ve had. Whichever print is better, l’ll keep for the next copy to be made from. The bad print will get thrown away.
I keep on doing this. Making a copy of my best print over and over again. Every now and then, the copier is going to make errors that I actually want. My best print will start to look more and more like the Mona Lisa.
To visualize this in a computer. Instead of a copier, you have a drawing program. You can feed it an image to start with, and it tries to make a copy with imperfections. But the imperfections aren’t random. They’re calculated by a formula.
You can use two randomly generated formulas and have them both draw a picture. One of those formulas is going to give better results than the other. Throw away the worse formula and make a copy of the good one with a few random changes.
You’ve now got two formulas you can test against one another again and repeat the process. Slowly getting better and better formulas over time.
Now for the final bit, which makes the AI learning much different from me and my copy machine.
You don’t want a formula that’s only good at drawing one image. You want the AI to make all sorts of images, even images you or the AI has never seen before.
If you were to feed it the same starting image and compare it to the same ending image every time. It would only be good at that one specific image.
You could end up with a formula that totally ignores the starting image and draws the Mona Lisa no matter what starting image you feed it. Or, a formula that heavily relies on that starting image and makes absolute nonsense when given a random starting image (like my copier).
Instead, you gather up a few hundred, or thousand images of nature (or cars, or animals, whatever) and draw a starting image for each one. Each time you test the formulas, you give it one of those starting images, but mix it up with different starting images for each test.
You already know what you want those starting images to look like in the end so you can compare what the formula makes with the corresponding images you have.
The better formulas aren’t going to be formulas that rely on an expected starting image or try to give the same result every time. A better formula will be one that takes whatever starting image you give it a does a good job at turning it into an image of nature (or cars, or animals, whatever).
Latest Answers