A study of feature selection and transfer learning in Electroencephalography based Brain Computer Interface
2017-02-27T00:05:51Z (GMT) by
Brain Computer Interface (BCI) is an interface system that allow direct communication pathway from brain to external devices bypassing conventional neuromuscular channel. The main motivation of such systems is to restore several lost abilities of individuals suffering from severe motor impairment; thus improving their quality of life. Electroencephalography (EEG) is by far the most widely used non-invasive neuro-imaging modality, owing to its practicality in real-world application. To date, session-to-session transferability and subject transferability are two of the major issues in BCI. The first objective of this research is to investigate the effect of channel selection on EEG data recorded with modalities with high electrode density setting. It is shown that variants of genetic algorithm can be used to study the optimal tradeoff between classification accuracy of a BCI system and the number of selected channels. Although the results from this stage shows that significant reduction in EEG electrodes can be achieved at expense of minor degradation in classification accuracy, the results also confirms the subject dependency and the non-stationarity of EEG signals. The non-stationarity of EEG signal is often associated with the constant changes in human mental state and this has impeded the efficiency in session to session transfer. The effect of non-stationarity is magnified in small sample setting where the small amount of training data available is not representative enough to derive a reliable feature extraction and classification model. The second research objective is to implement and evaluate Renyi entropy and fractal dimension algorithms as alternative feature extraction methods for motor imagery data. Renyi entropy proposed as an alternative feature extraction method has shown comparable performance with the best performing variant of the-state-of-the-art common spatial pattern (CSP) in various BCI settings. As compared to original CSP, Renyi entropy is less affected by the non-stationarity of EEG signals in small sample setting. In addition, the features extracted using Renyi entropy shows better subject transferability. The final objective of this research is to implement adaptation techniques to improve both inter-session and inter-subject transferability of EEG data. For subject-to-subject transfer, adaptation algorithms such as covariate shift can be used to weight the importance of each training sample, thus providing a more suitable learning model. As for session-to-session transfer, inter-session adaptation with covariate shift and inter-trial adaption with global mean updating and sequentially updated classification model is shown to enhance the performance of a BCI system.