Discipline: Data Science
Subcategory: Mathematics and Statistics
Session: 4
Room: Senate
Emily Diegel - Embry Riddle Aeronautical University
Co-Author(s): Rhiannon Hicks, Embry Riddle Aeronautical University, Daytona Beach FLMax Prilutsky, San Diego State University, San Diego CARachel Swan, Embry Riddle Aeronautical University, Daytona Beach FL
Neural networks are an emerging topic in the computer science industry due to their high versatility and efficiency with large data sets. Funded by the National Science Foundation, Embry-Riddle Aeronautical University is partnered with the Nevada National Security Site on the project, Ensemble Deep Learning, through the Research Experience for Undergraduates 2022 summer program. The Nevada National Security Site is looking to improve deep learning techniques that are used to analyze radiographic images of small-scale nuclear test explosions in order to ensure that the United States nuclear stockpile remains safe, reliable, and secure. Neural networks provide the tools to analyze these images due to the versatility of the model and their ability to analyze 3D data. However, neural networks are often referred to as a “black box algorithm” due to the underlying process during the training of the model. The explanation for a prediction is unable to be traced, therefore poses a concern on how robust and reliable the prediction model is. Building ensemble neural networks allows for the analysis of the error bars of the prediction model. The project’s objective is to determine the comparative differences between the predictive ability of each individual convolutional neural network versus the ensemble neural network. Additionally, we explored how to use the ensemble model as a method of uncertainty quantification. Adding an ensemble reduced the mean squared error from 0.003524 to 0.00268 and increased the R^2 value from 0.54251 to 0.65247 compared to an individual network. Overall, ensemble neural networks outperform singular networks and demonstrate areas of uncertainty and robustness in the model. For future research we plan to explore more ways of manipulating the training data to demonstrate further uncertainty and to improve the performance of the model.
Funder Acknowledgement(s): National Science Foundation
Faculty Advisor: Dr. Mihhail Berezovski, berezovm@erau.edu
Role: During my time on the project, I played a versatile role in managing the day-to-day research of the team while giving us an overall direction to go in. My individual accomplishments during this time included choosing an effective dataset, building the architecture of the neural network, implementing the training of the network, and analyzing the results to make the necessary corrections. Once the neural network was finalized, I investigated techniques for quantifying uncertainty and implemented the ensemble neural network.