Summary
International Symposium on Nonlinear Theory and Its Applications
2016
Session Number:A2L-D
Session:
Number:A2L-D-3
Multilayer Perceptron Including Different Amplitude Random Noise
Chihiro Ikuta, Yoko Uwate, Yoshifumi Nishio,
pp.-
Publication Date:2016/11/27
Online ISSN:2188-5079
Summary:
In this study, we introduce different amplitude random noise to a multilayer perceptron (MLP). We divide the neurons in hidden layer to some groups according to value of neuron output. Each neuron group is input the different amplitude random noise. When a neuron group has a large output, this neuron receives a large amplitude noise. When a neuron group has a small output, this neuro receives a small amplitude noise. The group member is dynamically changed during the learning because the neuron output changes with the learning. By simulations, we confirm that the proposed MLP performance is better than the standard MLP and MLP having simple noise input method. Moreover, we show the parameter dependency of the proposed MLP by changing the number of groups and the noise amplitude.