Semester

Fall

Date of Graduation

2019

Document Type

Dissertation

Degree Type

PhD

College

Eberly College of Arts and Sciences

Department

Mathematics

Committee Chair

Marjorie Darrah

Committee Co-Chair

Hany Ammar

Committee Member

Adam Halasz

Committee Member

Harvey Diamond

Committee Member

Jessica Deshler

Abstract

Artificial Neural Networks are extremely useful machine learning tools. They are used for many purposes, such as prediction, classification, pattern recognition, etc. Although neural networks have been used for decades, they are still often not completely understood or trusted, especially in safety and mission critical situations. Typically, neural networks are trained on data sets that are representative of what needs to be learned. Sometimes training sets are constructed in order to train the neural network in a certain way, in order to embed appropriate knowledge. The purpose of this research is to determine if there is another method that can be used to embed specific knowledge in a neural network before training and if this improves the performance of a neural network.

This research develops and tests a new method of embedding pre-knowledge into the Dynamic Cell Structure (DCS) neural network. The DCS is a type of self-organizing map neural network that has been used for many purposes, including classification. In the research presented here, the method used for embedding pre-knowledge into the neural network is to start by converting the knowledge to a set of IF/THEN rules, that can be easily understood and/or validated by a human expert. Once the rules are constructed and validated, then they are converted to a beginning neural network structure. This allows pre-knowledge to be embedded before training the neural network. This conversion and embedding process is called Rule Insertion.

In order to determine whether this process improves performance, the neural network was trained with and without pre-knowledge embedded. After the training, the neural network structure was again converted to rules, Rule Extraction, and then the neural network accuracy and the rule accuracy were computed. Also, the agreement between the neural network and the extracted rules was computed.

The findings of this research show that using Rule Insertion to embed pre-knowledge into a DCS neural network can increase the accuracy of the neural network. An expert can create the rules to be embedded and can also examine and validate the rules extracted to give more confidence in what the neural network has learned during training. The extracted rules are also a refinement of the inserted rules, meaning the neural network was able to improve upon the expert knowledge based on the data presented.

Download

Share

COinS