Yu-Ri Lee

A data partitioning method for increasing ensemble diversity of an eSVM-based P300 speller

Abstract A P300 speller is a device for typing words by analysing the electroencephalogram (EEG) caused by visual stimuli. Among classifying methods used for the P300 speller, the ensemble of support vector machines (eSVM) is well known for achieving considerable classification accuracy. The eSVM is composed of linear support vector machines trained by each small part of the divided training data. To obtain an ensemble model with good accuracy, it is generally important that each classifier be as accurate and diverse as possible; diverse classifiers have different errors on a dataset. However, the conventional eSVM considers only an accuracy viewpoint of an individual classifier by clustering the homogeneous training data with similar noisy components. With such a viewpoint of diversity, we propose a dataset manipulation method to divide a training dataset into several groups with different characteristics for training each classifier. We reveal that the distance between a letter on which a subject is concentrating, and an intensified line on a visual keyboard, can generate EEG signals with different characteristics in a P300 speller. Based on this property, we partition the training data into groups with the same distance. If each individual SVM is trained using each of these groups, the trained classifiers have the increased diversity. The experimental results of a P300 speller show that the proposed eSVM with higher diversity improves the letter typing speed of the P300 speller. Specifically, the proposed method shows an average of 70% accuracy (verbal communication with the Language Support Program is possible at that level) by repeating the dataset for a single letter only four times.

A reduced-complexity P300 speller based on an ensemble of SVMs

A brain computer interface (BCI) system is to control a computer using bio-signals measured in brain. A P300 speller is one of electroencephalogram (EEG)-based BCI systems. The speller is to display target characters which are what a subject wants to enter by detecting P300 wave. To detect the wave, a lot of EEG signals were averaged over the whole signals to increase the signal-to-noise ratio and the support vector machine (SVM) was applied to a P300 speller to separate EEG signals with P300 wave and without P300 wave in previous works. In current classifier topics, there are some methods to average some classifiers for performance improvement. An ensemble of SVMs is one of them but it has enormous computational complexity. To overcome this computational burden, we propose a P300 speller with preprocessing of channel selection and non-target data reduction. In conclusion, the calculation speed becomes higher than conventional method but, as a feature dimension decreases in channel selection part of the proposed method, the accuracy of the proposed method is lowered in both subjects.

AN ENHANCED EEG-BASED P300 SPELLER USING THE KERNEL ICA

A brain computer interface (BCI) system is to control a computer using bio-signals measured in brain. A P300 speller is one of electroencephalogram (EEG)-based BCI systems. The speller is to display target characters which are what a subject wants to enter. P300 wave, which is the most positive peak 260-410ms in an EEG signal after stimulus onset, is used as a control signal of the speller. The P300 wave has been separated using a blind source separation method in the existing P300 spellers. However, the conventional methods could not separate a source signal with Gaussian distribution from a set of mixed signals. To overcome this problem, we apply a kernel independent component analysis algorithm to P300 speller. The algorithm can successfully extract P300 component from a mixed signal even when it has source signals with nearly Gaussian distribution. In conclusion, the proposed P300 speller has 100% accuracy with less training signals and finds a target character more quickly than the conventional method.