A second-order learning algorithm for multilayer networks based on block Hessian matrix.

Abstract

This article proposes a new second-order learning algorithm for training the multilayer perceptron (MLP) networks. The proposed algorithm is a revised Newton's method. A forward-backward propagation scheme is first proposed for network computation of the Hessian matrix, H, of the output error function of the MLP. A block Hessian matrix, H(b), is then defined to approximate and simplify H. Several lemmas and theorems are proved to uncover the important properties of H and H(b), and verify the good approximation of H(b) to H; H(b) preserves the major properties of H. The theoretic analysis leads to the development of an efficient way for computing the inverse of H(b) recursively. In the proposed second-order learning algorithm, the least squares estimation technique is adopted to further lessen the local minimum problems. The proposed algorithm overcomes not only the drawbacks of the standard backpropagation algorithm (i.e. slow asymptotic convergence rate, bad controllability of convergence accuracy, local minimum problems, and high sensitivity to learning constant), but also the shortcomings of normal Newton's method used on the MLP, such as the lack of network implementation of H, ill representability of the diagonal terms of H, the heavy computation load of the inverse of H, and the requirement of a good initial estimate of the solution (weights). Several example problems are used to demonstrate the efficiency of the proposed learning algorithm. Extensive performance (convergence rate and accuracy) comparisons of the proposed algorithm with other learning schemes (including the standard backpropagation algorithm) are also made.