TESTING THE QUALITY OF NEURAL NETWORK ERROR CORRECTION FOR CALCULATING THE STANDARD DEVIATION OF SMALL SAMPLES OF BIOMETRIC DATA

• Published in: Измерение, мониторинг, управление, контроль no. 4
• Main Authors: Volchikhin, V.I., Ivanov, A.I., Malygina, E.A., Kachalin, S.V., Polkovnikova, S.A.
• Format: Journal Article
• Language: English
• Published: Penza State University Publishing House 01-01-2022
• Subjects: analysis of small samples; multilayer network of artificial neurons; prediction of errors in calculating the standard deviation
• ISSN: 2307-5538
• DOI: 10.21685/2307-5538-2021-4-8

Background. The statistical analysis of small samples in 16 experiments using the standard deviation estimation is considered. The aim of the work is the neural network prediction of errors in calculating standard deviations on small samples of biometric data. Materials and methods. Multi-layer networks of artificial neurons were used to predict the values of errors in calculating standard deviations. Deep neural network learning algorithms are well known. The main problem for their implementation is usually to obtain sufficiently large training samples. The novelty of the approach lies in the fact that for the problem under consideration, an automatic machine for forming training samples with different values of errors in estimating the standard deviation is used. Results. The created neural network error corrector reduces the error interval for calculating the standard deviation by 22.7 % for samples of 16 experiments. At the same time, the problem is revealed, which consists in the need to perform long-term training of multilayer neural networks for each new volume of samples. Conclusion. The analysis of the results obtained in the course of the study showed that neural network error correctors can increase the reliability of statistical estimates and other points. At the same time, neural network predictors of errors in calculating mathematical expectations and correlation coefficients can be created. It is assumed that the process of improving confidence will be monotonous and in one or two years it will be possible to reduce the uncertainty interval of calculations by an additional 20 % by using networks of 15 or 20 layers of neurons.