How Machines Learn From Examples

Pretend you want to teach a puppy to sit. You say sit, you gently help it sit, and you give it a treat. You do this many times. After a while, the puppy hears the word and sits on its own.

Computers learn in a surprisingly similar way. We show them tons of examples. We tell them when they are right and when they are wrong. Over time, they get better and better.

Teach a computer to spot cats

1. 1Show it 10,000 photos, each labeled cat or not cat
2. 2Let it make guesses
3. 3Tell it every time it gets a guess wrong
4. 4Let it adjust its brain a tiny bit after each mistake
5. 5Repeat until it gets most guesses right

After all that practice, the computer has built its own little cat detector. It did not memorize every photo. It learned the shape of cat-ness: pointy ears, whiskers, a certain kind of tail.

What can go wrong

• If all the example cats are orange, it may think black cats are not cats
• If someone labels dogs as cats by mistake, it will get confused
• If it only sees inside photos, it may not know outdoor cats

“Practice makes progress, not perfect, especially for computers.”

The big idea: computers learn by looking at lots of examples and fixing their mistakes over and over. The better the examples, the better the computer gets.