Smorodin A. Methods of neural networks training on the basis of nonlinear dynamics.

Dissertation solves an actual scientific problem important for the field of neural network learning. It contains new scientifically based results: - improved gradient descent methods based on nonlinear dynamics positions, which allowed to increase the convergence rate of gradient optimization algorithms during neural network training; - for the first time, a method for searching for parameters of improved gradient descent methods based on the positions of nonlinear dynamics on the basis of solving optimization problems on the class of complex polynomials was developed, which made it possible to ensure the effective implementation of the developed methods; - further development of methods of training neural networks on the basis of improved gradient descent methods, which made it possible to reduce the learning time of neural networks in solving segmentation and classification problems.