Structural-Parametric Synthesis of Convolution Neural Networks

Abstract

Structural-parametric synthesis of convolutional neural networks for solving the problem of image processing is considered. Classification of types of layers that form convolutional neural networks with the definition of their functional purpose and mathematical models that describe them are given. The procedure of weight coefficients initialization is considered in detail. Methods of overfitting avoidance of convolutional neural networks are analyzed. It has been shown that dropout has the greatest prospect. The definition of the most significant parameters of convolutional neural networks by the criterion of their effectiveness was carried out as a result of the experiment on a combined network consisting of a convolutional neural network, a classifier and a deconvolutional neural network, that allows not only to recognize the elements of the image, but also to mark the recognition elements on it. For the experiment, it was used database MNIST (database of samples of digits handwritten writing). The plan of the numerical experiment design with the purpose of convolutional neural networks parameters determination that most influence on the structural-parametric synthesis problem solution results is given. As an algorithm of optimization, it is proposed a genetic algorithm for the realization of which the chromosome size and structure is determined. The type of genetic algorithm operators and the size of the population are determined. An example of an optimal convolutional neural network constructing according to the criterion of accuracy is given.