Feature selection for improved classification

The authors apply the feature-selection technique of K. Fukunaga and W. Koontz (1970), an extension of the Karhunen-Loeve transformation, to spoken letter recognition. Feedforward networks trained for letter-pair discrimination with the new features showed up to 37% reduction in classifier error rate relative to networks trained with spectral coefficients. This performance increase was accompanied by a 91% reduction in feature dimension. For three-letter discrimination, the new features performed comparably to spectral coefficients, with a 90% reduction in feature dimension.<<ETX>>