That approach is almost exactly that of Deep Learning / stacked Restricted Boltzmann Machines.