Comparison Of Performance Between Traditional Back-propagation Network And Spiking Neural Network

Classification is one of the important fields in Artificial Intelligence. There are lots ofclassification model which can be applied to this field. The famous one is the traditionalMultilayer Feed-Forward Neural Network. With back-propagation learning, it is known as anoutstanding classifier to distinguish categories according to the desired output from the giveninputs. The traditional neural network performs well for classification problems in manyfields such as medical, science, engineering and business. Nevertheless, this traditionalback-propagation neural network still cannot serve all kind of data successfully, especiallythe temporal data.In recent decade years, the artificial neural networks have been developed continuously.To serve temporal data, a new generation of neural network has been established, which iscalled “Spiking Neural Network”, which simulates the biological neurons closer than thetraditional neural network does. The learning algorithm of spiking neural network is derivedfrom the traditional back-propagation, which is called as SpikeProp to take spike signals intoaccount. Though it has been trialed theoretically with the XOR problem, it has not yet beenwidely trialed with the real life data what the traditional neural network had tested. Therefore,comparing the performance between Traditional Neural Network and Spiking NeuralNetwork is explored in this thesis for general purpose use for real world data in classification.The performances for both neural networks are measured by the sum squared error andthe standard deviation. Experiments on10data sets have been trialed with traditionalback-propagation and SpikeProp algorithm, in which the error results from both algorithmshave been compared. The consequence shows that the traditional back-propagation is moresuitable for data sets with binary output than the spiking neural network; however, the spikingneural network is more preferable with more complex non-linear data sets than the traditionalback-propagation.