A Study on Visualizing Feature Extracted from Deep Restricted Boltzmann Machine using PCA

Abstract

P300 speller is a system that allows users to input letters using only electroencephalogram (EEG). A component called P300 is used to interpret the EEG in P300 speller. In order to achieve high performance in P300 speller, achieving high performance of P300 detection is essential. However, EEG waveforms are strongly dependent on the conditions of subject and/or environment, so it is not easy to detect P300 precisely. In this study, deep neural network using restricted boltzmann machine, which became famous by its high performance, is used to detect P300. It is expected that it also shows high performance for complex EEG waveforms. The experimental result shows that deep neural network was able to detect P300 better than the existing method (stepwise linear discriminant analysis). Furthermore, this study refers to the learned feature by deep restricted boltzmann machine. We can see that deep restricted boltzmann machine learns the feature extracted from the EEG waveforms correctly to detect P300 which led to the high performance.