Abstract. The paper shows that coefficients of correlation for biometric data are of considerable uncertainty if used for small test samples. This fact prevents from using them for machine learning (setting) of classical quadratic forms and Bayesian networks. The method of symmetrizing correlations is proposed to be used. It is proved that the requirements to the volume of biometric data are lower in this case. As a consequence, setting (teaching) of quadratic forms and maximum likelihood Bayesian networks become a much more stable problem. This is equivalent to the multiple reduction of requirements to the size of the training sample for “own” samples.

Keywords: biometric identification, symmetrization of correlations, machine learning based on small test samples.