Find the ceiling

1 · Find what changed

Beginner

Why did the same network fall from 0.97 to 0.50?

The planets of the second test were harder to see The network was too small for real data The false alarms of the second test look like planets

Ceiling of the AUC when a fraction f of the false alarms are blends, indistinguishable from planets: AUC = 1 − f/2. With f = 0.421, it is 0.789.

Intermediate

In another survey, 30% of the false alarms are blends. What is the best AUC a detector reading the light curve can reach?

0.70 0.85 0.15

Imagine a test in which every false alarm is a blend.

Expert

What AUC would a detector reach, even if perfect on everything else?

0.5, the level of chance 0, every pair ranked the wrong way 0.789, as in the article

Expert

Which change could lift a detector above the blend ceiling?

A deeper network trained on more simulated curves A second retraining on real false alarms Adding colour, position or star-motion data

—