Adrian Silvescu Fourier Neural Networks (Article)外语英文电子书.pdfVIP

Fourier Neural Networks Adrian Silvescu Artificial Intelligence Research Group Department of Computer Science Iowa State University, Ames, IA 50010 Abstract A new kind of neuron model that has a Fourier-like IN/OUT function is introduced. The model is discussed in a general theoretical framework and some completeness theorems are presented. Current experimental results show that the new model outperforms by a large margin both in representational power and convergence speed the classical mathematical model of neuron based on weighted sum of inputs filtered by a nonlinear function. The new model is also appealing from a neurophysiological point of view because it produces a more realistic representation by considering the inputs as oscillations. 1 Introduction The first mathematical model of a neuron was proposed by McCulloch Pitts[1943]. The underlying idea that this model tries to capture is that the response func- tion of a neuron is a weighted sum of its inputs filtered through a nonlinear function: = ( + ). Much progress has been done in the field of neural networks since that time but this idea still remained a very fundamental one. Although the model of the computational unit(neuron) per se is simple, neural networks are powerful com- puters, higher levels of complexity being achieved by connecting many neurons together. In this paper we try to propose more general and powerful models for the neuron as a computational unit. There are may motivations for this investiga- tion. One of them is the fact that although the power of computers increased quite a lot since 1943 we are still not able to simulate and train but toy-size neural networks. So