Abstract
In this note, we shed light on the physical meaning for the
backpropagated error used by the backpropagation training algorithm.
Essentially, for a given scalar output of the neural network, its
backpropagated error is a linear apportionment of the error at it, in
proportion to the linear gain between the outputs of neurons and the
output according to a linearised-systems-view of the network. For
multiple outputs, superposition provides the total/nett backpropagated
error at the outputs of neurons.
Subsequently, we present some elementary statistical analysis for
backpropagated errors in the network.