The blame assignment
Backpropagation is how a network learns which weights to change. After the forward pass produces a prediction, the loss measures how wrong it was. Backprop then works backward to find how much each weight contributed to that error.
- Forward computes the prediction and loss.
- Backward computes a gradient for every weight.
- Update nudges each weight to reduce the loss.
The intuition
Think of the error at the output as a signal that flows backward. Each weight receives a share of the blame in proportion to how much it influenced the result. A weight that strongly pushed the output in the wrong direction gets a large correction.
Why it is efficient
Computing gradients one weight at a time would be hopeless for millions of parameters. Backprop reuses intermediate results from the forward pass, so it finds every gradient in a single sweep backward. This efficiency is what makes training deep networks practical.
Key idea
Backpropagation flows the output error backward to assign each weight a share of the blame, computing every gradient in one efficient sweep so the network can update and improve.