Dropout Regularization in Deep Neural Networks Used in Atomic Simulations

Sandjakoska, Lj. (2018) Dropout Regularization in Deep Neural Networks Used in Atomic Simulations. “St Kliment Ohridski” University - Bitola, Faculty of Information and Communication Technologies - Bitola, Republic of Macedonia, pp. 52-55. ISBN 978-9989-870-80-4

Full text not available from this repository.
Official URL: https://aiitconference.org/2018/proceedings

Abstract

Recently deep learning has received significant attention in the research community. That implies the fast development of improved deep learning concepts, which can be applied in variety of domains. In most of the domains the problem of regularization arise. This paper aims to give a breakthrough in understanding the importance of dropout, as an adaptive regularizer that can control the overfitting in deep neural networks. Several aspects of the dropout regularization are discussed. We try to solve the problem of correlation between the adjacent units and propose a new method for regularization of deep neural network designed for atomic simulations, actually for predicting of the energies of small molecules. Also, a theoretical framework of the proposed method is given, followed by the experiments on molecular dynamics dataset. The main contribution of this paper is improving on state-of-the-art results of energy prediction using the equilibrium energies as well as molecular dynamics trajectory on benchmark dataset.

Item Type: Book
Subjects: Scientific Fields (Frascati) > Engineering and Technology > Electrical engineering, electronic engineering,information engineering
Divisions: Faculty of Information and Communication Technologies
Depositing User: Mrs Natasha Tabakovska
Date Deposited: 01 Mar 2019 23:04
Last Modified: 13 May 2019 08:48
URI: http://eprints.uklo.edu.mk/id/eprint/1491

Actions (login required)

View Item View Item