It computes gradients layer by layer so neural networks can learn.

Backpropagation applies the chain rule to calculate how each weight contributed to loss, then an optimizer updates weights.