Total least squares approach for fast learning in multilayer neural networks

Classical methods for training feedforward neural networks are characterized by a number of shortcomings, first of all the slow rate of convergence and the occurrence of local minima. In this paper a new learning algorithm is presented as a faster alternative to the backpropagation method. The algorithm is based on the solution of a linearized system for each layer of the network performed by a block total least squares technique. Simulation results are reported showing the high convergence speed of the new algorithm and its high degree of accuracy