METHODS OF BAYESIAN MATHEMATICAL STATISTICS IN TRAINING A CONVOLUTIONAL NEURAL NETWORK BASED ON A REGULARIZING BAYESIAN APPROACH

UDC 004.8
DOI: 10.36871/2618-9976.2022.09.003

Authors

Svetlana V. Prokopchina
Professor, Doctor of Technical Sciences, Financial University under the Government of the Russian Federation, Moscow, Russia

Abstract

One of the main stages of neural network development is the stage of neural network training and the development of a training dataset. The article proposes a new approach to training a Bayesian convolutional network with a probabilisticstatistical approach to development.
Dataset it is necessary to determine the probabilistic distribution of features or input signals of the neural network . Even for large samples, this is quite a difficult task. But most often such volumes of experimental information are not available in practical applied problems. The samples are either small or unique experiments giving single sample values. Such information situations refer to situations with significant information uncertainty. The article suggests using Bayesian mathematical statistics methods to determine the analytical type of distribution under conditions of significant uncertainty. An example of using this approach to prepare the training stage of the BCNN neural network is given.

Keywords

Bayesian mathematical statistics, Convolutional neural networks, Distribution law