Description of Algorithm for BCI Competition IV Dataset I (Y2008)

Team members: Zhang Haihong, Ang Kai Keng, Guan Cuntai, Wang Chuanchu, Chin Zheng Yang
Affiliation: Institute for Infocomm Research, Singapore

The algorithm consists of two parts

A. Machine learning of a robust estimator of user-state from EEG

In this part, we explore important spatio-spectral-temporal features that correlates with the three states: motor imagery class 1, motor imagery class 2, and the rest state.

A multi-model method is used in modelling the rest state EEG, where the rest epochs (short time windows) are clustered after applying PCA or ICA. Discriminative spatio-spectral-time features that differentiates each cluster of the rest state and the the motor imagery classes are learned through a filter-bank based classification method which learns common spatial patterns on each filter band and then applies mutual information based algorithm to select the features. A radial basis function based neural network is constructed to learn the mapping from the features to the desired outputs (i.e. -1, 0 or 1).

We use n-fold cross-validation to tune the parameters.

Note that no channel-selection is performed.

B. Processing evaluation data using the learned estimator

This part is straightforward. It processes the evaluation data sample by sample. At each sample, it extracts the features and maps them to the final outputs in the range of [-1 1].

Besides, the learning and processing are carried out in 100Hz. For the submission, the results are interpolated to 1000Hz.