How does AI learn?
Earlier you met a machine whose rules were written by hand. Write one rule, it spots one thing. Anything you didn't write down, it misses.
A learning machine is different. Nobody hands it the answer. It gets things wrong and fixes itself.
How can that work? It's far simpler than it sounds.
All you need is how far off you are
Picture a target. And one value you can turn. Turning it slides your answer left or right.
That value is called a . It sets how much notice to take of one signal. You nudge it a little at a time, like turning a dial.
How do you know which way to turn? Just look at how far off you are right now.
Turned it and got further off → turn the other way
Turned it and got closer → keep going that way
That's it. That really is it.
That "how far off" has a name: .
How many turns did it take? The graph underneath is what you just did. Way off at first, then smaller and smaller — that shape is learning.
The machine does the same thing
If you pressed Let the machine do it, it may have felt like a letdown. This is all it does.
Turn a little
Check: further off, or closer?
Go a bit more in the closer direction
Back to step 1
It isn't clever. It's just fast. You turn the weight once a second; the machine turns it thousands of times a second.
Now make it two weights
Press One more weight. It gets hard fast.
Turning one weight moves both answers. Line one up and the other slides away.
How many of these weights sit inside a real AI? Not hundreds — millions. Some have far more.
Could a person turn them all? No. So the machine turns them.
Sources for this chapter
- 6 Samuel, "Some Studies in Machine Learning Using the Game of Checkers", IBM Journal of Research and Development 3(3) (1959)
- 7 Samuel, "Some Studies in Machine Learning Using the Game of Checkers. II — Recent Progress", IBM Journal of Research and Development 11(6) (1967)