Complex-Valued Feedforward Neural Networks Learning Without Backpropagation

Abstract

This paper presents an efficient learning algorithm for complex-valued feedforward neural networks with application to classification problems. It simplifies complex-valued neural networks learning by using the forward-only computation rather than traditional forward and backward computations. By incorporating the forward-only computation, the complex-valued Levenberg-Marquardt algorithm becomes more efficient. Comparison results of computation cost show that the proposed forward-only complex-valued learning algorithm can be faster than the traditional implementation of the Levenberg-Marquardt algorithm.