Robotic Art Creation: AI

I'm finally talking about something that's been around for a year or two, namely art using Artificial Intelligence.  The basic principle--as far as I can understand it--is to give a program lots of images to 'eat' and digest, after which you can ask it to make a picture of a certain kind.

Obviously, it isn't enough to just shove a pile of images down it's throat.  The images have to be categorized.  For instance, if you're giving it geometric shapes, you've got to tell it: these are triangles, these are squares, and do on.  (This is an explanation for complete laymen; if you're an AI insider, you can stop reading now!)  So after digesting enough shapes, you would have succeeded in teaching it what the names of these shapes are.

This was back in the old days; modern AI Programs can be taught--what is the most exciting word for the guys who wrangle these programs--a girl!  (For some obscure reason, all the images of girls these programs are familiar with are those of girls with very ample chests, so naturally if one of these programs is requested to supply a picture of a girl, it's going to be one of a woman that appears to have been breast feeding for a while.)

The amazing thing about these programs is how they are able to combine different elements together.  Spitting out an image of a girl is not going to be a problem for a program that has access to a file with a few dozen pictures of various girls; similarly for an image of a tree, if it has a database of a few score of trees.  But you can see that it isn't simple for it to make a picture of a girl and a tree in the same picture!  The designer of the program must give it rules about how to combine images together in a satisfactory way, and they have certainly done this; this is why AI art is such an immense achievement.  Clearly, people have worked on the problem for a long while, and it isn't all the achievement of a single individual.

A while ago, I posted an image of a face I wanted to use as a selfie.  In the interests of complete disclosure, I made it clear that the image wasn't exactly me.  How did I pull it off?  Well, there's a free online app that combines faces, to produce a single face.  You give it two or more jpegs of faces, and it somehow averages them.  How does it do that?  It distills each face into a sheaf of numbers.  (For the moment, let's imagine that these are measurements, e.g. height of the nose, etc.)  Then, when it's given two faces, it just averages the numbers together.  Just as it can boil a face down into a vector of numbers, so it can "reverse-boil" the numbers back into a face, and that's what it did.  Now, I don't know exactly how all this is done; in fact, it's quite possible that some of the numbers aren't lengths, but angles, or ratios.  As you can imagine, averaging ratios might not work as easily as averaging simple lengths.  I'm just making sure we don't oversimplify the problem, and the solution.

Kay.