Conference paper Open Access

On the importance of deep learning regularization techniques in knowledge discovery

Ljubinka Sandjakoska; Atanas Hristov; Ana Madevska Bogdanova

Nowadays, in the era of complex data, the knowledge discovery process became one of the key challenges in the science. The evolution of the technologies imply evolution of the techniques for dealing with the data. Deep neural networks, as advanced concepts became very popular and can be viewed as tool for improvement of knowledge discovery processes. A motivation for this paper is generalization ability of deep neural network. In an attempt to better understand and to solve the problem of generalization of deep neural networks, we study several regularization techniques. Different regularization techniques, as a solution of overfitting problem, are discussed. The impact of regularization on knowledge discovery process is in the focus of this paper. In order to illustrate the effect of regularization in knowledge discovery, a case study is presented. The case study refers to discovering unknown relationships between molecules in atomic simulation. We propose a dropout method for regularization deep neural network for molecular dynamics simulations. In this paper we show that discovering high level concepts in data, during knowledge discovery, is possible with efficient training of regularized deep neural networks.

All versions This version
Views 109109
Downloads 8686
Data volume 41.5 MB41.5 MB
Unique views 9595
Unique downloads 6868


Cite as