Abstract
It is shown that large neural networks allow solving tasks that cannot classical quadratic forms in linear algebra. Thus the assessment of output entropy of neural network converters biometrics code is possible. The assessment of high-dimensional entropy is based on the symmetrization of the problem of the correlation of biometric data. Entropy of low dimension and high-dimensional entropy are differently connected with equally correlated data. For low-dimensional transformations only short-sighted algorithms, which not capable to bypass local extrema of quality are effective. The algorithms constructed on the accounting of multidimensional entropy are far-sighted, they don't see local extrema.