Application of entropic measures of stochastic dependence in pattern recognition

Abstract

It is shown that the new entropic measures of stochastic dependence among m variates recently introduced by Watanabe can be usefully employed in the solution of the feature extraction problem of pattern recognition. In particular it is shown that for minimizing the correlations among the components of the random vector in the feature space, in the Gaussian distribution case, the transformation matrix should be formed by using the m eigenvectors corresponding to the m largest eigenvalues of the correlation matrix, rather than the corresponding eigenvalues of the covariance matrix.