Weight limiting, weight quantisation and generalisation in multi-layer perceptrons

If a multilayer perceptron (MLP) is to be implemented on fixed point hardware then the robustness of the structure to weight quantisation is important. Most work on MLP performance totally neglects this issue and it is only addressed after a network has been trained. It is shown that both generalisation performance and robustness to weight quantisation can be improved by including explicit weight-range limiting into the MLP training procedure. This is illustrated by results of simulations on a speech recognition problem