My research focuses on Brain Computer Interface. Related fields include electroencephalography signal processing, high dimension data analysis, feature extraction and classification, and machine learning. The goal is to accurately translate human electroencephalography signals into useful commands for controlling personal computers and assistive devices.
Electroencephalography: Brain waves (electroencephalograms, EEG) are emitted constantly and can be picked up by sensitive electrodes on the surface of the scalp. Brain waves consist of groups of neurons firing with certain timing and locations. When engaged in certain mental tasks, such as imagination of moving limbs, memorizing numbers, or looking for an object of interest, the brain consistently emits certain kinds of brain waves depending on the mental task. We are hoping to use this consistency to build a computer software to decode the intention of a person based on the brain waves.
We analyze time series EEG and look for differences between mental tasks (classes). EEG is typically 1 to 35 Hz. Our lab samples EEG at 200 Hz and sometimes 1000 Hz. There can be anywhere from 16 to 64 electrodes in exploratory EEG experiments and 6 to 8 electrodes when area of interest has been located. These mean that in one second, 200 time points x 8 electrodes = 1600 data points are acquired. Typically, the time window useful for discriminating between different mental tasks is only 0.5 seconds wide. Still, it is a length-800 vector for every trial. The dimension of data dwarfs the number of trials. In laymen's term, there are more parameters than equations. We have two solutions to reduce the input dimension: Time binning and classwise principal component analysis (CPCA). Time binning merges multiple time points into one time point by averaging. It is essentially down sampling without the problem of aliasing. CPCA finds projection axes sorted from high variance to low variance of each class and between each pair of classes. It is ideal if classes are normally distributed. We keep principle axes whose eigenvalues are above mean. CPCA typically reduces input dimension to between 30 to 50 dimensions. This may or may not be good enough for classification, depending on the nature of the mental tasks. We usually go one step further and reduce to 1 to 5 feature dimensions by information discriminant analysis (IDA). IDA is a technique that maximizes certain objective functions based on information theory.
Brain Computer Interface in Communication and Control, 3 Oct 2009