Backpropagation: how a network learns by small corrections

In the last lesson we chose weights by hand. Real networks can have millions of weights, so nobody sets them manually. Instead, the network starts with random weights and improves them through many small corrections. Backpropagation is the method that tells each weight which way to move.