TESTING LARGE NEURAL NETWORKS ON SMALL S 

Alexander I. Ivanov, Doctor of technical sciences, associate professor, consultant, Penza Research Electrotechnical Institute (9 Sovetskaya street, Penza, Russia), E-mail: ivan@pniei.penza.ru
Igor A. Kubasov, Doctor of technical sciences, associate professor, chief researcher of the FKU NPO "STIS" of the Ministry of Internal Affairs of Russia, professor of sub-department of information technologies, Academy of management of the Ministry of internal affairs of the Russian Federation (8 Zoi i Aleksandra Kosmodem'yanskikh street, Moscow, Russia), E-mail: igorak@list.ru
Alexander M. Samokutyaev, Commander, Yu. A. Gagarin Cosmonaut Research Testing Center (Star city, Moscow region, Russia), E-mail: a.samo@mail.ru 

004.838.2 

10.21685/2307-4205-2021-1-5 

Background. The paper investigated the problems of testing neural networks used to improve the reliability and quality of complex technical systems. The conditions under which rapid and correct testing of the quality of decisions made by large neural networks on small samples is possible have been revealed.
Materials and methods. The probability of errors of the first kind (erroneous rejection of recognition of the "Own" image) was estimated based on testing without reducing the test sample. It has been shown that for neural network solutions in the form of a binary code, the probability of errors of the second kind (erroneous adoption of the "Alien" image) can be estimated with a significant reduction in the volume of the test sample.
Results and conclusions. A logarithmic decrease in the volume of the test sample was revealed when moving from statistical analysis of ordinary codes to statistical analysis of Hamming distances between the image code "Own" and the image codes "Alien." The mathematical model of calculation of probabilities of errors of the second kind of trusted neural network application on small samples is presented. The need for further standardization of trusted applications of artificial intelligence is justified, allowing to increase the reliability and quality of complex technical systems. 

artificial intelligence, neural networks, trusted neural network applications, test sample size