Mission control has a new delivery: two hundred fresh snapshot cards, and not one of them labeled.
"Sky Sorter needs labeled examples to learn from," Raven says. "Somebody has to label these."
Rocket groans, then grabs a pencil. Within an hour, he has labeled forty cards.
Raven checks his stack and holds one up. "This is a comet with a faint tail. You wrote star."
Nova hovers over the pile. "Every label is a lesson," she says. "A wrong label teaches a wrong lesson."
Rocket sighs and starts double-checking his whole stack.
Machine learning needs tremendous amounts of training data.
That training data usually has to be supplied by people. Sometimes the machine gathers it itself.
For Sky Sorter, people take the snapshots, measure the features and write the labels.
The patterns a sorter learns can only be as good as the examples it learns from.
| Card | Brightness | Size | Streak? | Label Rocket wrote | Correct label |
|---|---|---|---|---|---|
| R1 | 140 | 2 | no | star | star |
| R2 | 220 | 6 | no | planet | planet |
| R3 | 110 | 9 | no | star | comet |
| R4 | 190 | 1 | yes | satellite streak | satellite streak |
| R5 | 200 | 5 | no | planet | planet |
These cards are invented for our story.
Before training, careful teams check their data.
They look for missing numbers, impossible numbers and labels that do not fit.
A brightness of 300 is impossible on our 0 to 255 scale, so that card needs a second look.
| Card problem | Needs fixing? |
|---|---|
| Brightness 300 | ? |
| Size 4, labeled planet | ? |
| Size left blank | ? |
| Streak yes, labeled star | ? |
Great work! Tomorrow we learn why some labeled cards must be kept secret from the sorter.