The Study For A Self-Organization Mapping Network Based On Integrate-and-Fire Neurons

Nowadays, the artificial simulation of the behavioural characteristics of real neurons is one of the most important researches in neural computation. Although Self-Organizing Feature Mapping (SOM) has been widely studied both theoretically and practically, traditional SOM generally adopts the over-idealistic transfer functions, whose properties differ greatly from real neurons'characteristics. This dissertation discusses Intergrate-and-Fire (IF) artificial neuron model of high similarity to the neuron, and a new SOM model which uses IF neurons as nodes is founded.This paper begins with the research process of the Artificial Neural Networks (ANN), sketches the structure of real neuron, contrasts the four most frequently used artificial neuron model and clarifies the reason for establishing the new SOM with IF neurons.The most fundamental part of the article does a lot of researches about IF model on the basis of the elemental form and major characteristics of IF neurons. First, the membrane conductance form of IF model is introduced, and the effects of synaptic connection types and intensity on membrane potential is studied. Second, the paper analyzes the necessary conditions of globally synchronized action potential process, proves the reasonable time steps of the IF model experiment,and simplifies the IF model as the new transfer function to meet the demands of IFSOM. By doing these work, the thesis builds a new SOM network based on IF neuron and an topological structure of the SOM. Additionally, the dissertation picks one node out to describe the operation of the transfer function, whose results still have the physiological characteristics of real neurons.This dissertation concludes itself with the learning algorithm and procedures of the new-established IFSOM. With an example and experiment, the dissertation finds that the clustering function of the new-established IFSOM is superior to traditional SOMs, thus improving the traditional SOM. Finally, we generalize the advantages and disadvantages of the IFSOM network discussed in this dissertation.