12.5 Backpropagation

Backpropagation is how a network learns from its mistakes. After a forward pass, the error between the prediction and the true answer is measured, then propagated backward through the layers using the chain rule to compute how much each weight contributed to the error. An optimiser then nudges the weights to reduce that error.

Imagine a student in Addis taking a mock exam, seeing which questions they got wrong, and reviewing exactly those weak topics before the next test. Backpropagation tells the network which 'weak weights' to fix and by how much.