Aslarzanagh, Seyed Aliakbar Mousavi
(2013)
Classification Of P300 Signals In Brain-Computer Interface Using Neural Networks With Adjustable Activation Functions.
Masters thesis, Universiti Sains Malaysia.
Abstract
Brain-Computer Interface (BCI) employs brain’s Electroencephalograms (EEG) signals
and Event-related potentials (ERP) such as P300 to provide a direct communication between
human brain and computer. P300 speller application is a BCI that finds the location of target
character using P300 signals. This application tries to classify brain‘s P300 signals to find
the correct character from character board. P300 speller paradigm has two main classification
problems. The first problem is the detection of P300 signals from EEG data (classification of
P300 signals). Detection of P300 signals is a challenging task due to presence of noise and
artifacts in EEG data. The second problem is to correctly recognize the target characters based
on P300 signals. Detecting P300 signals is equivalent to detection of a character by a user who
was looking about 300 milliseconds before the signal detection. This study aims to classify
P300 signals with higher accuracy and recognize the characters with lower character trials by
using neural networks with adjustable activation function. The best neural networks model is
obtained by conducting three experiments on three NN models which differ based on the activation
function in the hidden layers and three standard classifiers. The performance of the best
NN model and its classifiers also compared with other classification techniques such as ESVM,
CNN and LDA in BCI. The results shows that neural network model NN3 with MoreletWavelet
activation function and multi-classifier MultiNC have obtained highest accuracy in P300 classification
and character recognition. It also shows that Sensitivity of P300 classification is better
describing the ranking of NN models and classifiers in character recognition problem.
Actions (login required)
 |
View Item |
