The one idea

Learning is guess → measure error → nudge knobs. Repeat.

A neural network starts out random and wrong. It makes a prediction, measures how wrong it was, adjusts its internal dials a tiny amount, and does it again — millions of times.