Backpropagation and unsupervised learning in linear networks