Chapter 06 / 17
How does AI actually work?
Many AI systems learn patterns from examples during training, then use what they learned to make predictions about new input.
Learning from examples
Imagine showing a system many labelled pictures of cats and dogs. It adjusts its internal settings to get better at guessing which label fits each picture.
Testing the result
A new picture tests whether the learned pattern works beyond the training examples. People need to check performance on different examples, including ones that are difficult or unusual.
Sam trains a card sorter
Sam made a pretend AI with picture cards of fruit. He labelled apples and oranges, then asked his cousin to hide a few cards for a test. At first, Sam's pretend system used colour alone, so it mistook a red apple for an orange. Oops.
They added more clues, such as shape, and tested new cards again. Sam learned that examples help a system find patterns, while fresh tests show whether those patterns work beyond the examples it has already seen.
A real-life example
A team training an animal-photo tool can label example pictures, teach the system from them, then test it on different pictures. The test helps reveal mistakes the examples did not show.
Test your own sorter
- Draw or find pictures of several kinds of leaves.
- Sort a few by one clue, such as shape.
- Ask a friend to add pictures you have not seen.
- Check where your sorting rule works or needs help.