Sending the error backward through a neural network to work out how to fix each weight.
Backpropagation is the core method for training neural networks with many layers. The network gives an answer and you measure the error. Then the error is passed backward, from the last layer toward the first, working out how much each weight contributed to the mistake and adjusting it by that much.
With one layer it's easy to see which numbers to blame; with many layers, the ones in the middle are hard to judge. Backpropagation solves this with the chain rule from calculus. Earlier work described it in the 1970s, and a 1986 paper by Rumelhart, Hinton and Williams made it widely known. Today's large language models are still trained this way.
Think of it like this
When a group project gets a low mark, you trace back from the final presentation through each step to see where things went wrong and by how much. Big contributors to the mistake get fixed a lot; small ones get fixed a little.
For example
If an AI calls a cat photo a "dog", backpropagation nudges millions of numbers that fed into that mistake, all at once.
Records about this
Source-checked moments from AI history.
Showed that hidden layers learn useful internal representations for a task on their own, making the training of multilayer neural networks practical
Regarded as the first description, predating the 1986 rediscovery. The reference point in the debate over priority for backpropagation