A new network’s connections are random, so its first answers are mostly wrong.
Signals flow through the layers. Strong weights pass them on.
The error is traced back, and every weight moves a little to shrink it.
With every example the answers improve, until the weights capture the pattern.