The Gini differential test for neural network analysis of small biometric data samples

Abstract

Backgroud. Training of artificial neural networks in accordance with the algorithms for recognizing biometric patterns described in GOST R 52633.5-2011 should be performed on small samples. The research considers samples consisting of 16 and 64 experiments. Materials and methods. Shown for samples of 16 experiments, the chi-square test of 1900 gives unacceptable error probabilities. The classic 1941 Gini test gives error probabilities 29% higher than the chi-square test. Results and conclusions. The use of the proposed new differential version of the Gini test allows to obtain results in samples of 16 experiments approximately 9 times better than those of the chi-square test in samples of the same size. A variant of neural network use of classical and newly created criteria on their basis is considered. At the same time, the neural network use of 9 new statistical criteria makes it possible to reduce the probability of errors of the first and second types to a value of 0.031. If the worst-case test is replaced with the Gini differential test in this group, the probability of errors can be reduced to a value of 0.027.

Keywords

• fast neural network training algorithms
• small samples
• statistical criteria for testing the normality and uniformity hypothesis
• binary neurocriteria
• artificial neurons
• improving the accuracy of estimates by eliminating code redundancy