Building AI by letting the computer learn rules from examples instead of being told.
Machine learning means you don't hand the computer the rules. You show it lots of examples and let it work the rules out. To catch spam, instead of listing "spammy" words one by one, you show it tens of thousands of spam and normal emails.
While looking at the examples, the computer keeps nudging the numbers inside it (its parameters) so it makes fewer mistakes. That process is called training. There are several styles: supervised learning gives it the right answers, reinforcement learning gives it rewards.
Think of it like this
Nobody learns to ride a bike by memorizing a manual. You wobble, fall, adjust and try again. Machine learning works the same way: practice and correction instead of a rulebook.
For example
A music app suggesting songs like the ones you keep replaying is machine learning.
Common mix-ups
Mix-upMachine learning is a program with a huge number of rules written in by people.
ActuallyWhat people decide is the big frame: which examples to show and what counts as a mistake. The actual decision rules come out as numbers, so often even the makers can't read them line by line.
Mix-upWhen a machine "learns", it understands the way a person does.
ActuallyThe machine just picks up patterns that tend to show up alongside the right answer. If most of the wolf photos have snow in them, it can learn to say "wolf" because of the white background rather than the animal.
Try it in class
For teachers at the start of a lesson — or for students working on their own.
- Question
- Which jobs suit "writing the rules" better, and which suit "showing examples"?
- Activity
- With a partner, try to describe "pick out the photos of dogs" with rules alone (ear shape, fur colour…). Count how many lines it takes before the rules break down, then talk about why showing lots of example photos is easier. Finally, open the 1958 Perceptron record (E0005) and check how long learning machines have been around.
Records about this
Source-checked moments from AI history.
The first learning procedure to update weights from errors. Psychological Review 65(6) 386–408