Geometrical initialization, parametrization and control of multilayer perceptrons: application to function approximation

This paper proposes a new method to reduce training time for neural nets used as function approximators. This method relies on a geometrical control of multilayer perceptrons (MLP). The geometrical initialization gives better starting points for the learning process, and so the geometrical parametrization achieves a more stable convergence. During the learning process, a dynamic geometrical control helps to avoid local minima. Finally, simulation results are presented, showing a drastic reduction in training time and an increase in convergence rate