Overfitting

A network can score almost perfectly on the 8,000 photos it trained on and badly on the 2,000 it has never seen.

That gap has a name: .

It is like the student from the previous slide. They learned the practice paper so narrowly that a new version of the exam still catches them out.

A network can latch onto details that happen to sort its training photos, such as a background, a watermark, or one photographer's framing.

That is why the 2,000 photos are held back. They reveal whether the learned pattern still works when the examples change.