A comparative study of the Kohonen and multiedit neural net learning algorithms

This paper presents a comparative evaluation of the multiedit/condensing and Kohonen neural net learning algorithms using a speaker-independent speech recognition problem as a test vehicle. Both approaches attempt to cover the subspaces associated with respective pattern classes by a small number of reference vectors for subsequent nearest neighbour classification of unknown patterns. Several important design issues are addressed such as feature selection, use of alternative distance metrics, learning strategy, the form of adaptation function and the number of reference vectors. Results obtained using the k-nearest neighbour rule are also presented for comparison