Neural Network Classification of Word Evoked Neuromagnetic Brain Activity

Abstract

The brain-physiological signatures of words are modulatedby their psycholinguistic and physical properties. Thefine-grained differences in complex spatio-temporalpatterns of a single word induced brain response may bedetected using unsupervised neuronal networks. Objective ofthis study was to motivate and explore an architecture of aKohonen net and its performance, even when physicalstimulus properties are kept constant over the classes. Weinvestigated 16 words from four lexico-semantic classes.The items from the four classes were matched for wordlength and frequency. A Kohonen net was trained on the datarecorded from a single subject. After learning, the networkperformed above chance on new testing data. The resultsobtained suggest that the research on single trialrecognition of brain responses is feasible and a rich fieldto explore.