Open AccessTraining algorithms for convolutional neural networks
Author(s) -
Arsentiy Igorevich Bredikhin
Publication year - 2019
Publication title -
vestnik ûgorskogo gosudarstvennogo universiteta
Language(s) - English
Resource type - Journals
eISSN - 2078-9114
pISSN - 1816-9228
DOI - 10.17816/byusu20190141-54
Subject(s) - mnist database , computer science , convolutional neural network , correctness , artificial intelligence , artificial neural network , algorithm , task (project management) , sample (material) , machine learning , pattern recognition (psychology) , chemistry , management , chromatography , economics
In this article we consider one of the most used classes of neural networks convolutional neural networks (hereinafter CNN). In particular, the areas of their application, algorithms of signal propagation by CNN and CNN training are described and the methods of CNN functioning algorithms implementation in MATLAB programming language are given. The article presents the results of research on the effectiveness of the CNN learning algorithm in solving classification problems with its help. In the course of these studies, such a characteristic of the neural network as the dynamics of the network error values depending on the learning rate is considered, and the correctness of the algorithm of learning convolutional neural network is checked. In this case, the problem of handwritten digits recognition on the MNIST sample is used as a classification task.