Research On Complex-Valued Feedforward Neural Networks Learning Without Backpropagation

In recent years,as an important branch of the research field of artificial neural networks,complex-valued neural networks are becoming more and more popular with researchers.People popularize the network structure and learning theory of neural networks from real fields to complex fields,and complex numerical signals are gradually used to solve practical problems.Feedforward neural network has the characteristics of simple network structure,concise training process,excellent approximation performance and fast convergence speed,so the feedforward neural network has been widely used.In this paper,the feedforward full complex-valued neural network and its algorithm are studied,and a feedforward flull complex-valued neural network with no backward propagation is proposed to improve the operation efficiency of the network structure.First of all,we propose a gradient descent algorithm for the real and imaginary type complex-valued feedforward neural networks.The algorithm is extended from the traditional split complex forward neural network.It can be seen from the experimental analysis that the complex-valued feedforward neural network algorithm has powerful computing power and better generalization ability.At the same time,through experimental comparison and analysis,we can see the advantages of complex-valued neural network without backward propagation.Then,this thesis presents an efficient Levenberg-Marquardt learning algorithm for the real and imaginary type complex-valued feedforward neural networks and fully complex-valued neural networks,which simplifies eomplex-valued neural networks training by using the forward-only computation instead of traditional forward and backward computation.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-ouly complex-valued computation can be faster than the traditional implementation of the Levenberg-Marquardt algorithm.We apply the proposed algorithm to the real data set classification in the UCI public database.The experimental results prove the effectiveness of the proposed method in this thesis.