A network begins with a structure: neurons arranged in layers, connected by weights.

Training changes those weights. Examples set the task. Backpropagation calculates how small weight changes would affect the error. Gradient descent uses those calculations to update the weights.

Across many corrections, the hidden layers form useful intermediate representations. Success means those learned patterns also work on relevant examples outside the training set.

The structure may look the same after training. The weights do not.

So what remains when training ends? Where does the learned behaviour live, and what happens when the network is used on a new input?

That is the question the next chapter will answer.