Date of Award


Document Type


Degree Name

Bachelor of Science



First Advisor

Dr. Art Charlesworth


This thesis does not assume the reader is familiar with artificial neural networks. However, to keep the thesis concise, it assumes the reader is familiar with the standard Machine Learning concepts of training set, validation set, and test set [1]. Their usage is intended to help ensure that the Machine Learning system can generalize its training from input examples used during its training to “similar” kinds of examples never used during its training.

The concept of a Convolutional Neural Network (CNN) is one of the most successful computational concepts today for solving image classification problems. However, CNNs are difficult and time- consuming to train. Such training requires automatically adjusting parameters to find a choice of parameter values that achieves good classification accuracy. In order to achieve good accuracy, a CNN must sometimes have several million parameters, which can cause the automatic training to require several days on even very powerful current hardware. Moreover, designing a successful CNN can require a wise choice of a dozen or more hyper-parameters, a concept further explained in Section 4. This thesis investigates the relationship between a certain hyper-parameter and the CNN’s resulting accuracy using that hyper-parameter on different applications. Our main research question is stated in Section 4. It uses terminology that needs to be explained first.