Raven spreads 16 labeled photos across the floor. "Now we look for patterns," she says.
Rocket points. "Every plant photo has leaves!"
"What do you notice about the two tools with no handle?" Raven asks.
Nova hovers low. "Here is a hint," she says. "Count each feature for each label. Then compare."
Machine learning finds patterns in data.
One way to find a pattern is to count.
Count how often each feature shows up for each label.
If a feature shows up for one label again and again, that is a pattern.
Here is the crew's training data: 16 labeled photos, described by their features.
| Label | How many photos | Their features |
|---|---|---|
| tool | 4 | shiny metal, a handle, straight edges |
| tool | 2 | shiny metal, no handle, straight edges |
| plant | 5 | green, leaves |
| rock | 3 | gray, lumpy |
| rock | 2 | shiny and wet, gray, lumpy |
| Feature | tool | plant | rock |
|---|---|---|---|
| a handle | 4 | 0 | 0 |
| straight edges | 6 | 0 | 0 |
| shiny | 6 | 0 | 2 |
| leaves | 0 | 5 | 0 |
| lumpy | 0 | 0 | 5 |
Rocket's old rule looked for shiny things. It mixed up tools and wet rocks.
The data shows a better clue. Straight edges only show up in tool photos.
The training data taught the crew a pattern they missed.
| Statement | True or false? |
|---|---|
| Straight edges show up only in tool photos here. | ? |
| Shiny photos are always tools. | ? |
| Counting features can reveal a pattern. | ? |
| Machine learning finds patterns in data. | ? |
Super pattern finding! Tomorrow you train your own paper model in the lab.