Development of a Deep Learning-Based Brain-Computer Interface for Visual Imagery Recognition

Electroencephalogram (EEG) signals are naturally chaotic and nonlinear. This, however, poses problems in the decoding of brain states for recognition and classification applications, particularly in brain-computer interfaces (BCIs). Recent BCI technologies are focused on the extraction of linear and nonlinear signal features. However, these systems do not necessarily capture the chaoticity of EEG signals, and hence, may affect the accuracy of classification. Furthermore, most recent EEG-based BCI systems involve nonlinear feature extraction; however, these systems require explicit and hand-selection of these features, and may therefore impose further errors. In this study, we investigate the use of deep learning techniques that involve inherent and embedded feature selection and extraction in their hidden layers, and hence, they do not require explicit and human-intervened selection of features. The performance of different deep learning models were compared to those of the traditional classifiers. Results show that deep learning-based BCI systems performed significantly better in comparison to the conventional classifiers. This BCI system can be extended in different applications, such as in recognition of imagined phonemes, words, phrases and sentences.