Ensemble of Neural Network Conditional Random Fields for Self-Paced Brain Computer Interfaces
Hossein Bashashati, Rabab Kreidieh Ward · Advances in Science Technology and Engineering Systems Journal · 2017
Classification of EEG signals in self-paced Brain Computer Interfaces (BCI) is an extremely challenging task.The main difficulty stems from the fact that start time of a control task is not defined.Therefore it is imperative to exploit the characteristics of the EEG data to the extent possible.In sensory motor self-paced BCIs, while performing the mental task, the user's brain goes through several well-defined internal state changes.Applying appropriate classifiers that can capture these state changes and exploit the temporal correlation in EEG data can enhance the performance of the BCI.In this paper, we propose an ensemble learning approach for self-paced BCIs.We use Bayesian optimization to train several different classifiers on different parts of the BCI hyperparameter space.We call each of these classifiers Neural Network Conditional Random Field (NNCRF).NNCRF is a combination of a neural network and conditional random field (CRF).As in the standard CRF, NNCRF is able to model the correlation between adjacent EEG samples.However, NNCRF can also model the nonlinear dependencies between the input and the output, which makes it more powerful than the standard CRF.We compare the performance of our algorithm to those of three popular sequence labeling algorithms (Hidden Markov Models, Hidden Markov Support Vector Machines and CRF), and to two classical classifiers (Logistic Regression and Support Vector Machines).The classifiers are compared for the two cases: when the ensemble learning approach is not used and when it is.The data used in our studies are those from the BCI competition IV and the SM2 dataset.We show that our algorithm is considerably superior to the other approaches in terms of the Area Under the Curve (AUC) of the BCI system.