Multichannel classification of brain-wave representations of language by perceptron-based models and independent component analysis

Electroencephalography (EEG) has been used since the 1920s. In the past few decades, research has been based on averaging trials to provide information on evoked response potentials. Here, I study both average and individual trials. The analysis is formulated as a classification task with supervised learning using multichannel EEG data. Category labels were assigned to each stimulus. Both visual and auditory stimuli were presented, including images, syllables, words and sentences---all related to natural language.; For both average and individual trial classifications, significant rates were obtained with p-values less than 10-10. These results serve as good evidence for the existence of a brain wave representation for words and sentences. By combining a multichannel multi-output perceptron with Independent Component Analysis, I was able to further improve the already significant single channel results, both strengthening the evidence and opening up possibilities for brain-computer interface applications. The details of various machine learning techniques, statistical evaluations and parameter interpretations were discussed in this manuscript.