Neural Network Training Via a Primal-Dual Interior Point Method for Linear Programming

We propose a new training algorithm for feedforward supervised neural networks based on the primal-dual interior point method for linear programming. Specifically we consider single layer networks where the error function is defined by the L₁-norm and the activation function of the output layer is linear. Because of the special structure of the problem, our problem is equivalent to solving a block-structured linear system of equations.