Computational Experience with Pseudoinversion-Based Training of Neural Networks Using Random Projection Matrices