EEG Model for Classifying Dominant Images in Binocular Rivalry

EE362/Psych221 Final Project - Winter 2009
Blair Bohannan and Steinunn Arnardottir

 
 
Background
Stimuli
Methods
Analysis
Results
Conclusions
Acknowledgements
References
Appendix
Contact
 
Home

Artifact Removal
Decomposition of sources is performed with Independent Component Analysis (ICA) to determine which sources are related to noise artifacts such as eye movements and find sources in the brain that are more correlated to the task. For this the Infomax ICA algorithm from Matlab EEGLab toolbox is used. Linear decomposition of channel data finds the most independent sources (as linear combinations of the channels).

Three formats of data are cut into trials:
- raw data
- ICA sources
- ICA sources with noise artifacts removed (sources highly correlated with eye movements)

 
   

Classification
To separate the sources, Principal Component Analysis (PCA) is used with K principal components. K is optimized between 1 and 200 in a nested cross-validation loop. For the feature reduction to a dimension below 200, a Single Value Decomposition (SVD) is used.

Linear Discriminant Classification (LDC) with tenfold validation is used for classifying sources. That is, cross-validation partitions the trials into 10 equal-sized sets. 90% of the trials are used for training and 10% for testing every time. This is repeated ten times for all the different parts of testing data, and the average of the 10 classification rates is an estimator of the mean classification rate.

To build a multi-class classifier from a two-class classification method, the one-against-all strategy is used. That is, several two-class problems are defined and run against each other. For all dimensions, we compare pair-rates, 4-class rates and 8-class rates.

The reason for choosing a linear classifier in this case, over a non-linear one, such as the Support Vector Machine (SVM), is the limited amount of data available. In this kind of applications, linear classifiers have shown to perform better than non-linear ones (Perreau Guimaraes 2007).