In this paper, a quadrics and BP (back propagation) based neural network model is proposed. with the division of the input vactors into several sub-vectors and the application of quadrics, the multi-linear spaces are separated well, so that the complex net connection between the input level and the implicated level is eliminated and physical meanings of the hidden neurons are clarified. compared with the traditional three-level BP neural network, the result of the experiment show that this improvement realizes faster and more stable convergence.