Rocket has been patching Sky Sorter all weekend. "Card E is a bright star, so I raised the brightness cut-off to 215."
Raven slides a new card across the console: brightness 205, size 6, no streak. "This is a planet. Your new rule calls it a star."
Rocket groans. "Every time I fix one card, another one breaks."
Nova drifts over the growing pile of patches. "You keep guessing the cut-off yourself," she says.
"What if the cards could show you where it belongs?"
Raven pulls out a fresh stack of cards, every one already labeled.
There are two big ways to give a computer a sorting job.
Way 1: people write the rules by hand. The computer follows them exactly.
Way 2: the computer learns from data. It finds patterns in many labeled examples.
The second way is called machine learning. It is a kind of statistical inference that finds patterns in data.
| Hand-written rules | Machine learning | |
|---|---|---|
| Who chooses the rules? | A person | The pattern found in the data |
| What does it need? | A person who knows a rule for every case | A great many labeled examples |
| How do you fix a mistake? | A person edits the rule | Improve the data, then learn again |
| Sky Sorter so far | Version 1, written by the crew | Coming this week |
Rocket only looked at a few cards at a time, so each patch fixed one card and broke another.
When a job has many tricky cases, hand-written rules can grow long and fragile.
Learning from data looks at many examples at once. That is its big strength.
| Statement | True or false? |
|---|---|
| A hand-written rule is chosen by a person. | ? |
| Machine learning finds patterns in data. | ? |
| Rocket's 215 cut-off fixed every card. | ? |
| Hand-written rules can grow long and fragile when there are many tricky cases. | ? |
Great start! Tomorrow we look closely at how hand-written rules work and where they come from.