Geometric structure of shallow neural networks and constructive ${\mathcal L}^2$ cost minimization