Common Nature of Learning Between Back-Propagation and Hopfield-Type Neural Networks for Generalized Matrix Inversion With Simplified Models