Well, that’s pretty much like training a neural network, except that the integration of the matrix coefficients happens not from the immediate output data, but from data which was output by the network millions of epochs ago.