New principles of finding and removing elements of mathematical model for reducing computational and time complexity

The original principle of removing elements of a mathematical model based on its parametric identification of neural network is proposed in the paper. The essence of proposed method is to find a functional subset with less variable results and higher accuracy than for the initial functional set of the model. It allows reducing the computational and time complexity of the applications built on the model. The comparison with dropout technique shows the 1, 1 decreasing of root mean squared error. In addition, reducing the complexity allows increasing the accuracy of neural network models. Therefore, reducing the number of parameters is an essential step in data preprocessing used in almost all modern systems. However, known methods of reducing the dimension depend on the problem area, making it impossible to use them in ensemble models.