A feedforward neural network with one hidden layer is applied to the problem of brainstem auditory evoked potential classification. Network performances were tested separately both on subject-dependent samples (drawn from the same subjects from which the training set was derived) and on subject-independent samples (drawn from subjects from which no data was included in the training set), and compared. The results indicate that increasing the training set size improves performance, and human selected training sets give better results than randomly selected sets. Different encoding schemes of signal representation yielded a wide range of correct recognition rates. A spectro-temporal representation was very successful in terms of generalization to novel input and for producing high recognition rates for both classification categories. Although the networks were overly complex and trained in the memorization mode, they displayed feature extracting and generalization capabilities.
|Original language||English (US)|
|Number of pages||11|
|State||Published - Dec 1 1992|
ASJC Scopus subject areas