Improvement of neural networks learning by feature extraction methods
2013
Kodors, S., Rezekne Higher Education Institution (Latvia)
This paper discusses a comparison of the feature extraction methods, if a classifier of recognition problem is the artificial neural network with the back-propagation algorithm. The feature extraction methods can improve the classification accuracy and minimize a size of an education dataset and a signal processing time. All these improvements are satisfied by the transformation of the recognizable signal and by the minimization of the signal size. There are two problems to measure these improvements: the improvement of the recognition can be only determined in the experiment, the second problem is that a measured structure of the artificial neural network can contain the unlimited number of the layers and the unlimited number of the perceptrons in every layer. Therefore there is need to argument the chosen parameters of the experiment. The goal of this work is to organize the experiment plan to compare the feature extraction methods. The paper contains the description of the structure of the artificial neural network, the dataset and the elements which are influenced by the feature extraction methods.
显示更多 [+] 显示较少 [-]