一种复值函数型连接神经网络

本文提出了一种复值函数型连接神经网络(CFLNN)结构,可以对复数域信号进行快速处理。函数型连接神经网络通过对输入模式预先进行非线性扩展,增强了输入信号的模式表达,从而可以大为简化网络结构,降低计算复杂度。本文将函数型连接神经网络推广到了复值情况并给出了基于梯度下降的学习方法。计算复杂度分析显示本方法具有结构简单,计算量低的优点。最后,将本方法运用到对复值非线性系统的辩识问题中,仿真实验表明本CFLNN性能与传统复值前馈神经网络相近或更优。

Nonlinear System Identification Using Complex-valued Chebyshev Functional Link Neural Networks

A complex-valued functional link neural networks (CFLNN) for the purpose of fast complex domain signal processing is proposed in this paper. A functional link neural network can expand its input pattern to eliminate the need of hidden layer without sacrifice its performance. Thus the network structure and the computational complexity can be remarkably reduced. In this paper, we employ this proposed complex-valued functional link neural network to the problem of nonlinear system identification and compare its performance with the conventional complex-valued feedforward neural network. The computational complexity analysis and simulation result indicate that the performance of CFLNN is similar or superior to that of a multiplayer perceptron, accompanied with a much simpler structure and lower computational complexity.