9How does it work?

It only learns what you show it

The machine in chapter 8 saw 18,000 pictures. This time you teach the machine.

Posong's cards go in Posong's box, Arong's cards in Arong's box.

Once they're sorted, press Posong. The machine gets it right. Even though it's a new picture that was never in a box.

It does one thing. It finds the three closest cards. If two of the three are Posong, it says Posong. The bigger side wins.

Those three stay on screen, so you can see exactly why it answered that way.

Now press a dashed friend. You never taught it that one.

It answered anyway, didn't it? It can't say I don't know. It only learned two boxes, so it has to pick one of them.

The machine isn't broken. That's just all it was shown.

A little deeper

For grown-ups, and for children who want more.

This classifier is k-nearest neighbours with k set to three: find the three most similar cards, take the majority. It is among the simplest methods in machine learning.

Simple was the point. Its reasoning shows on screen as it works. A neural network would do better but has a hard time showing you why. The card you ask about is a pose that was never put in a box — ask with a card you put in and you are only watching it recall what it memorised, so we did it this way to see whether it applies what it learned to a new picture.

When the cards lean one way, the answers lean with them. That is bias — not a fault in the machine but the ordinary consequence of what it was shown. Real cases where face recognition fails more often on certain ethnicities, or hiring filters score one gender down, have exactly this shape.

The dashed friend matters. With only two boxes learned, the machine picks one of the two, always. Nobody ever gave it the option of saying it does not know.