Training Data

Also known as: Examples, The Examples We Show It

Training data is the pile of examples you show a computer so it can learn something.

If you only ever showed the computer pictures of white dogs, then one day it sees a black dog and says: that is not a dog.

It is not being silly. It only knows what we showed it. We gave it a bad pile of examples, so it learned a wrong thing.

Good examples make a computer that is right a lot. Careless examples make a computer that is confidently wrong, which is worse, because it still sounds sure.

Color this: make a good pile. Color four dogs that all look different from each other. That pile teaches better than four dogs that look the same.

For the grown-up holding the crayon box:

This is where bias enters, and children understand it best as unfairness rather than statistics: a machine that only saw one kind of example treats everything else as wrong. Teaching this at six costs nothing and pays off for decades.