Backpropagation
Also known as: backprop
Backpropagation (backprop) is how a network learns. After the model makes a prediction, backprop:
- Measures how wrong it was (the loss).
- Works backward through every layer to figure out which weights contributed to the error.
- Nudges each weight a little in the direction that reduces the error.
Repeat that millions of times and the model converges toward good answers.
The big idea: training = prediction + backprop + nudge, looped billions of times. Everything else in ML is plumbing around this loop.