Geometric structure of Deep Learning networks and construction of global ${\mathcal L}^2$ minimizers