BAYESIAN APPROACH IN NEURAL NETWORKS

Abstract

With the development of Deep learning, neural networks have become very popular. But nowadays neural networks which are being used are based on standard approach of Statistics. The main purpose of this work is to present Bayesian approach, highlight the main differences between Bayesian and frequentist approaches, their principles The problem was set to show in which cases Bayesian neural networks can be more preferable. The purpose was achieved through the following stages. The main steps of neural networks are presented. After it the fundaments of Bayesian approach are described. As Bayesian approach is a common concept, not just an inference used in neural networks, initially author speaks about this approach generally and then tells about its usage in neural networks. After that the main advantages and limitations of Bayesian networks in Deep learning are spoken about: which problems they can solve. The main conclusion of this paper is that Bayesian approach could be used to avoid overfitting. However, it is essential to understand the specific conditions under which Bayesian neural networks are preferable.