Rocket wants to fix the night side problem fast. "Add one dark photo. Done!"
Raven counts the training pile out loud. "Forty bright photos and one dark one."
"What do you notice about those numbers?" she asks.
Nova floats between the two piles. "Is one dark example enough to learn from?" she asks.
Machine learning needs lots of examples to find a pattern.
One dark photo is not enough. The model needs many dark photos too.
The crew adds dark photos of rocks, plants and tools.
Then they test again on both bright and dark test cards.
| Training pile | Bright photos | Dark photos |
|---|---|---|
| Before | 40 | 1 |
| After | 40 | 30 |
| Test after the fix | Right | Tried |
|---|---|---|
| Bright photos | 9 | 10 |
| Dark photos | 8 | 10 |
Gap hunters read data tables and ask: what is missing or tiny?
A row with 0 or a very small number is a clue.
| Training set | Does it have a gap? |
|---|---|
| 20 rock, 20 plant, 0 tool photos | ? |
| 15 rock, 15 plant, 15 tool photos | ? |
| Photos from day side only | ? |
| Kid voices and grown-up voices, 25 of each | ? |
Excellent gap hunting! Tomorrow we think about fairness in everyday life.