The New Jersey Institute of Technology's
Electronic Theses & Dissertations Project

Title:	Characteristics of different deep neural networks and application of pre-trained model without transfer learning
Author:	Peng, Zhiqi
View Online:	njit-etd2017-109 (x, 37 pages ~ 1.4 MB pdf)
Department:	Department of Computer Science
Degree:	Master of Science
Program:	Computer Science
Document Type:	Thesis
Advisory Committee:	Wei, Zhi (Committee chair) Roshan, Usman W. (Committee member) Phan, Hai Nhat (Committee member)
Date:	2017-12
Keywords:	Deep neural networks Data simulations
Availability:	Unrestricted
Abstract:	Deep neural networks have been successful in many areas, some of them even surpass human performances. The goal of this thesis is using data simulations to present different characteristics of three deep neural networks: fully connected deep neural network, convolutional neural network, recurrent neural network, which will perform best when dealing with different feature patterns. By using these characteristics to design a deep neural network on top of an adopted pre-trained model with untrainable layers, achieved an averagely 11.1% improvement than a model with transfer learning method.