Strictly local backpropagation

In conventional implementations of the backpropagation method, the delta error terms which propagate the error back through the network are global (common) variables shared by multiple nodes. A modified implementation of the algorithm which permits this process to occur using only local values of all variables is presented. This innovation is beneficial for massively parallel implementations. It is only the network architecture that has been modified; the computational scheme is completely equivalent to the standard algorithm. This modification was made to address criticisms of the backpropagation model on grounds that it is not biologically plausible