CGP Grey explains how machines learn


#1

Originally published at: https://boingboing.net/2017/12/19/cgp-grey-explains-how-machines.html


#2

#3

If you know a bit about programming and are comfortable with high school mathematics, then I heartily recommend taking Andrew Ng’s Coursera course:

It’s one of the best courses I’ve taken online or off. Professor Ng is an excellent teacher and the course is very well designed.


#4

We know perfectly well how and why bots make decisions. The problem is that the reasoning is too complex for us to understand. It’s the same with human reasoning: if you ask them why they know something is a bee, the only real answer would be because it’s bee-shaped. Whatever else they tell you it’s just bullshit. You would recognize a bee in a black and white photo, or without wings. Deep inside your brain you have a pattern of what a bee should look like and anything close to it gets classified as a bee. And that human pattern is a lot more inscrutable to us as the pattern in a machine learning algorithm.

Also, we have had this kind of black-box algorithms for a long time. I mean, a linear regression does the same thing. Let’s say you have data about house sizes and price. A linear regression will tell you that prices go up with size, but will not tell you the reason why. We can guess a reason (people like bigger houses), because the it is simple. But as bots try to find more and more complex relationships, then, at some point, humans lose the ability to guess the reason.

That is also not how machines learn. They can be trained like that, what he described is a simple genetic algorithm. But it’s not efficient at all, so it’s not done often. Besides, genetic algorithms do have theoretical guarantees, despite what the video said.

What we actually do is a lot more like teaching. You start with a bot with random connections and test it on a bunch of examples. Then you tune the connections based on the answers it got wrong or right. And repeat the process for a couple thousand times.

He is right about training when the bot is trying to optimize something for humans. We are too weird, irrational and indecipherable, so all you can do is make random changes and just go with whatever works.


#5

Yay - I love CGP Grey. I wish he would do more great videos.


#6

This topic was automatically closed after 5 days. New replies are no longer allowed.