I saw a wonderful talk today about machine learning. Most of the time when people talk about machine learning they deal in abstractions. They write down some math, they wave their hands, they mutter vaguely about neural networks, and in general they say things that are completely mysterious to most of the populace.
But the talk today, by Chris Olah, was anything but mysterious. He pretty much laid it out for us, in terms that anybody could understand.
Essentially, machine learning algorithms are like Silly Putty. They take the space of all of the variables that go into whatever an algorithm is trying to recognize, and they stretch and distort that space in all sorts of interesting ways.
After all that distortion, whatever it is the algorithm is supposed to recognize ends up on one side of some plane, and everything else ends up on the other side. For example, if the machine learning algorithm is trying to recognize pictures with dogs in them, then after all the Silly Putty distortion, all the pictures containing dogs will end up on one side of the plane, and all of the pictures without dogs will end up on the other side.
Then it’s just a matter of using a mathematical knife to cut through that plane. On one side will be all the dog pictures, on the other side will be the non-dog pictures.
And that my friends, in a nutshell, is what machine learning is all about. I had no idea, until today, that Silly Putty could be so useful.