training, eval
A student who memorizes last year’s test looks brilliant until the questions move.
If you only practice one exam, you learn the exam. You do not learn the subject. Overfitting is that failure in a network. The weights latch onto quirks in the training clips — a specific camera smear, a painted logo on a local bus, the exact timing of one city’s left-turn arrow.
On the training set the scores look excellent. On a new street they fall over. That is why teams hold out data the model is not allowed to train on, and why a good lab result can still disappoint in another city.
You do not learn the subject.
Overfitting is also why “just add more data” is incomplete advice. More data from the same quirk teaches the quirk harder. You need new variation, or penalties that punish memorizing noise.
From the rider’s seat overfitting can look like superstition. The car slows at a harmless billboard that appeared in a bad training clip. Or it handles one parking garage perfectly and a similar garage poorly. The model did not understand garages. It remembered a garage.