Two-Stage Second Order Training in Feedforward Neural Networks

In this paper, we develop and demonstrate a new 2nd order two-stage algorithm called OWO-Newton. First, two-stage algorithms are motivated and the Gauss Newton input weight Hessian matrix is developed. Block coordinate descent is used to apply Newton’s algorithm alternately to the input and output weights. Its performance is comparable to Levenberg-Marquardt and it has the advantage of reduced computational complexity. The algorithm is shown to have a form of affine invariance.