Backpropagation is the core training method that makes deep neural networks practical. After a network makes a prediction, the algorithm measures the error, then works backward layer by layer to calculate how much each internal weight contributed to that error, and nudges every weight slightly to reduce it. Repeating this over many examples gradually turns a random network into an accurate one.
The method was popularized by David Rumelhart, Geoffrey Hinton, and Ronald Williams in their 1986 paper “Learning representations by back-propagating errors,” published in Nature, volume 323. They showed that hidden layers could learn useful internal representations on their own, directly answering the limitations of single-layer perceptrons.
This was the breakthrough that revived neural network research after years of stagnation, because it gave a concrete, efficient recipe for training networks with multiple layers. Essentially every modern deep learning system, from convolutional networks to large language models, is trained with a descendant of backpropagation.
Why business readers should care: backpropagation is the engine that converts raw data and computing power into a trained model. The cost, time, and hardware needed to “train an AI” are almost entirely the cost of running backpropagation at scale, which is why training frontier models is so expensive.