Neural Network Classification of Word Evoked Neuromagnetic Brain Activity

Abstract

The brain-physiological signatures of words are modulated by their psycholinguistic and physical properties. The fine-grained differences in complex spatio-temporal patterns of a single word induced brain response may, nevertheless, be detected using unsupervised neuronal networks. Objective of this study was to motivate and explore an architecture of a Kohonen net and its performance, even when physical stimulus properties are kept constant over the classes.

We investigated 16 words from four lexico-semantic classes. The items from the four classes were matched for word length and frequency. A Kohonen net was trained on the data recorded from a single subject. After learning, the network performed above chance on new testing data: In the recognition of the neuromagnetic signal from individual words its recognition rate was 28% above chance (Chi-square = 16.3, p<0.0001) and its accuracy was 44% above chance (Chi-square = 40.8, p<0.0001). The classification of brain responses into lexico-semantic classes was also unexpectedly high (recognition rate 16% above chance, Chi-square = 27.2, p<0.0001, accuracy 20% above chance, Chi-square = 42.0, p<0.0001).

Our results suggest that research on single trial recognition of brain responses is feasible and a rich field to explore.