2022 REU Summer Projects
Intermittency Analysis in Chaotic Systems
REU student: Ethan M. Kessler
Mentors: Dr. Trung (Tim) Le (NDSU) and Dr. Alvaro Valencia (University of Chile)
Project description: Complex physiological processes have commonly exhibited chaotic dynamics and intermittent phenomena. Such complex systems have been analyzed and decomposed into intermittently forced linear systems by means of the Hankel alternative view of Koopman (HAVOK) method. Furthermore, wavelet analysis methods have been used to further analyze the spectral and temporal properties of the chaotic intermittent bursting in systems such as apneic events in obstructive sleep apnea (OSA) patients. In this paper, we present further methods to analyze the intermittent bursting in these systems, using OSA as a case study to validate our results. We first extract Heart Rate Variability (HRV) features from the electrocardiogram (ECG) signals of the OSA patients. Then, we perform HAVOK analysis to decompose the system into a linear dynamics component, and an intermittent forcing component. Using predetermined apnea annotations, we confirm how the forcing term correlates to the apneic events. Next, we perform a continuous wavelet transform to obtain spectral and temporal information simultaneously. Finally, we address the main limitations of the previously explored methods. We develop an adaptive threshold for the forcing component to better correlate with the apneic events. We also combine the wavelet coefficients at multiple frequencies to capture broad-frequency-band intermittency. Results show that our methods improve the correlation for all patients, implying that filling these gaps in the previous research will help better control this advanced pathophysiological process for better preventative treatment. Our future work will attempt to tackle the main limitations of these methods, including a generalized solution for the case when apnea annotations are not provided.
Federated Learning for Internet of Medical Things
REU student: Taranatee (Tara) Khan
Mentors: Dr. Trung (Tim) Le (NDSU) and Dr. Alvaro Valencia (University of Chile)
Project description: Traditional, centralized machine-learning models have a multitude of limitations. The need to share data from remote clients to a main central server causes high latency and lengthy training times. Furthermore, the sharing of data poses a security risk. This is particularly worrisome for models including Internet of Medical Things devices, which contain sensitive data. Federated learning is a decentralized type of machine learning, where the learning mainly occurs at each client. A learning parameter is then sent from each client to a central server, where the local models are aggregated. This allows for improved training times and data security, by decreasing the risk of data interception. This project utilizes the N-BaIoT dataset, which features nine different household IoT devices undergoing both Mirai and BASHLITE types of attacks. The anomalies are detected using an autoencoder algorithm. The data is preprocessed using
min-max normalization for each feature and a threshold is then developed using the reconstruction mean square error of the autoencoder. If the reconstruction error is significant, the data’s equation surpasses the threshold and is classified as anomalous. Both centralized and federated learning models were trained using strictly benign data, and the results were compared. Results indicate that accuracy for federated learning and centralized machine learning are comparable. Federated allows for a significant decrease in training time, while also decreasing the risk of a privacy or data breach.
Collaboration in Research
REU student: Blaine Farrell
Mentors: Dr. Nita Yodo (NDSU) and Dr. Benjamin Herrmann (University of Chile)
Project description: This project is centered on the use of Academic Social Networks and Social Network Analysis to discover the impact of collaboration on the rank of authors in research and if there are disparities between fields of study. By utilizing Scholarly Big Data in a Microsoft Academic Graph, an analyzable dataset was formed to allow for a deep dive into what drives author success and the differences among sixteen fields of study. Success for one’s rank was pre-calculated by the Microsoft Academic Graph creator and determined by the determined probability of importance assigned to an entity. After creating a dataset, the correlations between given features and paper rank, author rank, journal rank, and conference rank were found to explore the importance of networks on an author’s rank or if any other patterns arose that may explain one’s rank or how collaboration can be defined in research. These findings were later broken down into each field of study to determine any differences that may arise or show a pattern. Although there were problems with the size of the data, complexity with having multiple entities (nodes), and lack of more descriptive features, these problems were avoided through the use of Social Networks and their analysis to provide a better description of each author node and to organize the data. After finding the most important factor to the success of a paper’s rank was the average of its authors’ rank, the use of collaboration networks proved worthwhile. Also, the use of a network contributed to an increased amount of meaningful features among author entities and allowed for a deeper look into the impact of relationships and connections that may result in a better individual author ranking. This project seeks to find what causes an author to have a higher probability of importance, the value of collaboration in determining this importance, and the future promotion of meaningful collaboration in academia.
Enzyme Classification with Embedding Methods
REU student: Ilya Tataurov
Mentors: Dr. Saeed Salem (NDSU) and Dr. Benjamin Herrmann (University of Chile)
Project description: Graph classification is vital in drug development, chemistry, and biology. In this project, fixed features from enzyme graphs were extracted. Furthermore, various machine learning models were trained on the best features and evaluate each model’s predictive power. The advantage of this approach is the speed and robustness of the predictions.
Drone Urban Scene Image Translation and Enhancements
REU student: TeVaughn Shaw
Mentors: Dr. Simone Ludwig (NDSU) and Dr. Mónica Zamora Zapata (University of Chile)
Project description: The emergence of deep learning in the safety and landing procedures of Unmanned Aerial Vehicles (UAVs, or “drones”) has generated valuable research topics over the years. This paper will consider the difference in generated urban scene UAV images based on Pix2Pix GAN and Real-ESRGAN frameworks. Computer vision plays a crucial role in the autonomous operations that go into UAV navigation technologies to ensure flight safety and avoid crowds or other surrounding obstacles. The experimentation results of the different generative adversarial networks (GANs) will be conducted on the Semantic Drone Dataset. They will exceed the typical generative drone image research by using enhancement techniques to produce realistic images.
Hardware-aware Neural Architecture Search
REU student: Evan Scully
Mentors: Dr. Uma Tida (NDSU) & Dr. Zamora Zapata (University of Chile)
Project description: Neural Architecture Search (NAS) is a novel field of Machine Learning and Deep Learning. A simplified description of NAS breaks down into a machine learning algorithm that chooses the best possible Neural Architecture or Neural Network possible for a given problem. With NAS, there are some considerations that need to be considered, such as the specific hardware that will be running the architecture. This is where Hardware-Aware Neural Architecture Search (HW-NAS) comes into play. In HW-NAS, the Search Strategy takes into account the capabilities of the hardware that the neural network will be run on. In a typical NAS, the network’s Search Strategy is only tasked with looking into the Search Space to find the best network possible via the Performance Estimator, while in a HW-NAS the Search Strategy also takes in the latency, energy consumption, etc. of the assumed hardware.
Defect Detection in Additive Manufacturing
REU student: Alexander Torres
Mentors: Dr. Nita Yodo (NDSU) and Dr. Benjamin Herrmann (University of Chile)
Project description: Additive manufacturing has grown in popularity in the past decade and has become the prominent way to quickly manufacture objects for both industrial and domestic use. Although Additive manufacturing, also known as 3D printing, has been quickly adopted in many areas of manufacturing, the area of defect and error detecting is still a problem than exist especially in lower cost fabricators. Current works toward detecting defects focus on computer vision and image recognition to detect defects in real time. Such defects can include warping, bed detachment, extrusion errors such as a clogged nozzle, depleted material, and over extrusion. This project explores the use of a convolutional neural networks to detect print failures in various datasets.
Image Recognition in the Beef World
REU student: Hailey Bixler
Mentors: Dr. Rex Sun (NDSU) and Dr. Ruben Fernandez (University of Chile)
Project description: Machine learning has the capability to turn guesses into informed decisions. This includes consumer purchase decisions. With a machine learning-based mobile app, customers can take a picture of a packaged steak to discover its cut and quality. The app uses machine learning to determine the cut of the beef and calculates aspects such as marbling and tenderness. The model, image recognition process, and user interface were already implemented, but the classifier had yet to be tested and the runtime averaged at 10.18 seconds per use case, which was more than most consumers would be willing to wait. These two areas were the focus of the project. The initial model was tested on four of its eight possible labels and determined to have an accuracy rate of 4%. By reducing the image size and therefore dimensionality, the runtime of the app was decreased to about 6.16 seconds. This was done without significantly altering the accuracy rate - when tested on the same four labels, the accuracy rate of the new model was about 3%. However, when tested on images from all eight possible labels, the accuracy rate turned out to be closer to 20%, which was better than random guessing. However, this is still not ideal. The low accuracy rate is likely caused by an improper training dataset, meaning that to achieve more accurate results, a new dataset must be collected and used to retrain the model.