Использование генеративно-состязательных нейронных сетей для расширения обучающих (тестовых) выборок

Abstract

The article is devoted to the formation of the initial data set for training neural networks. The main methods of formation and expansion of training (test) samples are presented, in particular with the help of augmentation, mathematical modeling methods, as well as generative-adversarial neural networks. The principle of operation of GAN networks is considered, the algorithm of operation of the generator and discriminator according to the minimax game method, the Wasserstein method and the "penalty gradients" method is described. A digital simulation of the GAN network was performed using a minimax model trained on the MNIST dataset. Conclusions are drawn regarding the possibility of using the developed generative-adversarial neural network in order to expand the bank of training (test) samples.