CONCEPT
Backpropagation
The universal training algorithm of modern artificial intelligence—the procedure that sends an error signal backward through every layer of a neural network and adjusts each connection by a hair in the direction that would have reduced the mistake, repeated until the network learns.
Backpropagation is the closest thing modern artificial intelligence has to a first principle. Every deployed large neural network—every language model, every image classifier, every speech system, every game-playing agent that learned its skill—was trained by computing an error and propagating it backward through the network to adjust the weights. The procedure was made practical for networks of neuron-like units by
David Rumelhart, Geoffrey Hinton, and Ronald Williams in a 1986 paper titled “Learning representations by back-propagating errors.” Its mathematics were the chain rule of calculus, old by 1986; its decisive contribution was demonstrating that the procedure actually worked to discover useful internal representations in hidden layers that no human had specified. Because everything a network can do was learned by this single blind procedure of error reduction, the intelligence of the system is real in its effects and blind in its origin—it knows nothing but the gradient of an error it is told