AI & Computer Science · Weekly AI Papers
Backpropagation: Learning Representations by Back-Propagating Errors (Rumelhart, Hinton & Williams)
LISTEN TO THIS EPISODE
Backpropagation lets a neural network adjust its internal connections using the errors in its outputs.
David Rumelhart, Geoffrey Hinton and Ronald Williams show how hidden units can learn useful representations in their 1986 Nature paper, Learning representations by back-propagating errors.
A forward pass computes the outputs, and a backward pass calculates how changes in each weight affect the error.
The discussion considers examples in which hidden units discover regularities in the task.
This is a foundational account of an artificial learning procedure, rather than a demonstration that the brain implements it in the same way.
Read about the original source: Learning representations by back-propagating errors
Part of Marginalia: books, papers, and ideas for learning and debate.