Method of automatic search for the structure and parameters of neural networks for solving information processing problems

Abstract

Neural networks are actively used in solving various applied problems of data analysis, processing and generation. When using them, one of the difficult stages is the selection of the structure and parameters of neural networks (the number and types of layers of neurons, activation functions, optimizers, and so on) that provide the greatest accuracy and, therefore, the success of solving the problem. Currently, this issue is being solved by analytical selection of the neural network architecture by a researcher or software developer. Existing automatic tools (AutoKeras, AutoGAN, AutoSklearn, DEvol and others) are not universal and functional enough. Therefore, within the framework of this work, a method of automatic search for the structure and parameters of neural networks of various types (multilayer dense, convolutional, generative-adversarial, autoencoders, and others) is considered for solving a wide class of problems. The formalization of the method and its main stages are presented. The approbation of the method is considered, which proves its effectiveness in relation to the analytical solution in the selection of the architecture of the neural network. A comparison of the method with existing analogues is carried out, its advantage is revealed in terms of the accuracy of the formed neural networks and the time to find a solution. The research results can be used to solve a large class of data processing problems for which it is required to automate the selection of the structure and parameters of a neural network.

Keywords

• neural networks
• machine learning
• optimization of the neural network structure
• data analysis and processing
• convolutional neural networks
• autoencoders
• generative-adversarial networks