Comparison of Gaussian process models for single-trial event-related potentials

In this work we present a comparative study of Gaussian process models for single-trial event-related potentials (ERPs) in electroencephalography (EEG) recordings. Our data comes from a motor task experiment where an ERP arises before the motor response of the participant to a stimulus. We consider models based on stationary and non-stationary kernel functions. The comparison is done based on two different criteria: model likelihood and model reaction time prediction. We show how models with high likelihoods do not necessarily perform well at predicting reaction time. The non-stationary kernel function achieved the best predictive performance.