A neural network stacks layers of simple units: each hidden unit draws a soft straight edge through the input space (tanh of a weighted sum), and the next layer mixes those edges into curved boundaries. Training nudges every weight downhill on the loss, with gradients from backpropagation.