10.31673/2412-9070.2020.050311

Abstract

The article considers an artificial neural network — a multilayer perceptron, as well as the Kohonen network, as tools for visual image recognition. The work is devoted to the analysis of the main functional tasks that can be solved by our chosen neural networks, with maximum efficiency and expediency. We will also consider these neural networks as tools for primary analysis of visual images for further use both for scientific purposes and for their substantive implementation in business areas where neural networks are most applicable and will bring maximum effect to ordinary users. One of the most common algorithms for learning multilayer neural networks is the error backpropagation algorithm. This algorithm calculates the gradient surface gradient vector. Then we move by some value in the direction of the vector (it will show us the direction of the fastest descent), where the error value will be less. This consistent progress will gradually minimize the error. Here there are difficulties with definition of size on which it is necessary to advance. If the size of the step is relatively large, it will lead to the fastest descent, but there is a chance to "jump" the desired point or go in the wrong direction, if the surface has a fairly complex shape. For example, if the surface is a narrow ravine with steep slopes, the algorithm will move very slowly, jumping from one slope to another. If the step size is small, it will lead to finding the most optimal direction, but can significantly increase the number of iterations. To achieve the most optimal result, the step size is taken in proportion to the steepness of the slope with some constant - the speed of learning. The choice of this constant is made experimentally and depends on the conditions of a particular problem. Having conducted the work, it should be noted that the multilayer perceptron is successfully used to solve various problems, in particular as effectively as possible for the problem of pattern recognition.