Skip to content
Exit presentation

How AI Learns I: From Rules to Data

Learning the examples instead of the pattern

Show the program every crate. Can it still be wrong?

Loading lab…

All twenty-four crates become examples.
Step 1 of 4
  1. All twenty-four crates become examples.
  2. Ask about a crate it holds: distance 0. It copies the word.
  3. Score on its own examples: 24 of 24.
  4. Score on the next morning's twenty: 14 of 20.

The rule

Yes. It can still be wrong on a crate it has not seen. A crate is held when it is one of the program's examples. Distance 0 means the same marked count and the same kilos.

A nearest-example program that holds a crate finds it at distance 0. So it copies that crate's word. That is automatic, not a skill. The only count that matters covers crates it has never seen.

  • A perfect score on its own examples proves nothing new.
Two scores, the same program
Measured onScore
Its own 24 examples24 of 24
The next morning's 2014 of 20

The examples hold both full-price and cut-price crates. The program still misses 20 − 14 = 6 of the next twenty.

Jump to lesson

This track

  1. 0 Start here
  2. 1 A rule a person wrote, and a pile of examples
  3. 2 Keep some crates back
  4. 3 Right out of how many
  5. 4 Split it differently
  6. 5 Learning the examples instead of the pattern
  7. 6 Ask more than one
  8. 7 The order you do it in
  9. 8 Write the rule yourself
  10. 9 The big picture
  11. 10 Track quiz

Other tracks