Algorithmic enhancements to a backpropagation interior point learning rule

In this paper, the authors employ a quadratic interior point method to backpropagation neural networks. The new quadratic backpropagation learning rule searches for a direction which minimizes the objective function in a neighborhood of the current weight vector. Numerical results on the parity problem show that the new learning rule is more than ten times faster than the standard backpropagation, and five times faster than the linear interior point learning rule developed earlier by the same authors (1995)