The station turns toward the dark side of its orbit. Rocket snaps new mission photos in the dim light.
He feeds them to the sorting model. It calls almost every photo a rock!
"But it scored 9 out of 10 last week," Rocket groans.
Raven looks at the old training cards. "What do you notice about these photos?"
Nova glows softly. "Look at the light in each one," she says.
Every training card the crew made was a bright photo.
The test cards were bright too. So the model scored well.
But the model never learned from a dark photo.
When dark photos came, it did not know what to do.
Training data that leaves out some kinds of examples is one-sided.
The missing kind of example is a gap.
| Photos | Right | Tried |
|---|---|---|
| Bright photos | 9 | 10 |
| Dark photos | 4 | 10 |
| Statement | True or false? |
|---|---|
| The training cards were all bright photos. | ? |
| The model learned from many dark photos. | ? |
| A gap is a kind of example that is missing. | ? |
| A good test score on bright photos proves it works on dark ones. | ? |
AI can help, and it can also cause harm.
If training data is one-sided, an AI can work worse for some people.
People who build AI should look for gaps and make it fair.
Sharp noticing! Tomorrow we learn how one-sided data can be unfair to people.