Designed in, not learned

That arrangement was chosen by people before training began.

Almost everything else here was discovered by training. Nobody told AlexNet what an edge was, or which measurements mattered. The weights found that themselves.

But sliding a small detector across the image was not discovered. A person decided it, before training started, and built it into the shape of the network.

The decision encodes a useful assumption about photographs: the same local pattern may matter in different places. A small curve can help identify an ear near the edge or the middle of a picture. Researchers built that reuse into the architecture instead of requiring separate weights at every position.

So a network is really two things. A structure somebody designed, and numbers that training filled in.

Keep the distinction. When you hear that a system learned something by itself, the useful question is what was left for it to learn, and what was decided in advance by whoever built the shape it learned inside.