This summer, the Department of Computer Science & Engineering welcomed 12 students from peer institutions to the Research Experiences for Undergraduates (REU) Program. The 10-week program focused on human-centered computing for social good in the areas of virtual reality, visualization, human-computer interaction, and human-robot interaction. Each project addressed at least one of the United Nations Sustainable Development Goals with students presenting their work at the Summer Undergraduate Research Expo (SURE) to cap off the summer.
Hosted by Evan Suma Rosenberg, the REU program guides undergraduate students through an entire research experience with weekly research seminars and workshops, and hands-on experience working in labs with graduate researchers. Department mentors included Zhu-Tian Chen, Victoria Interrante, Harmanpreet Kaur, Daniel Keefe, Rosenberg, Junaed Sattar, Qianwen Wang, and Lana Yarosh.
“One thing I think that I learned is that when it comes to research, there are a lot of places you can jump off from and you need to read the latest papers,” said Shautavia Hooper, an undergraduate from Carleton College. “There is so much information out there about embodiment and cybersickness and reading papers was a great way to get started and figure out what you find interesting.”
“I definitely enjoyed this program,” said Cynthia Maldonade. “I really enjoyed working with the people in the lab and my peers. The people I worked with were really nice and I thought, ‘Oh maybe I want to stay here.’ I really enjoyed it.”
Learn more about each student’s research project below!
—
Maryam Bacchus
Project: Understanding the impact of generative AI assistance in writing-based tasks
Advisor: Harmanpreet Kaur
Home Institution: Vassar College
Abstract: Following the public release of ChatGPT in late 2022, generative AI (genAI) tools have been increasingly embedded in a variety of knowledge work tasks (e.g., writing, coding, information retrieval). We focus on the impact of genAI on writing: individuals across disciplines are now using genAI for both creative and non-creative tasks such as writing articles, cover letters, and application forms. However, with the fast pace of technological advancement in this setting, we have yet to fully realize the impact of using genAI for these tasks. Therefore, we consider the following research questions: (1) how does the process and outcome of writing change when people use AI-assisted tools, and (2) to what extent does personality play a role in this behavior? In order to test these questions, we conducted a within-subjects pilot study where participants completed two writing tasks—with and without an AI-assistant—as well as a short exit interview. Our results speak to the changes in the writing process between the two conditions, attention to detail, and whether or not an individual's levels of conscientiousness and agreeableness connect to their performance.
Catherine Bao
Project: Enhancing LLM Accuracy in Self-Diagnosis: Strategies for Optimizing Performance With Accessible Toolkits
Advisor: Qianwen Wang
Home Institution: University of Utah
Abstract: With large language models now achieving nearly 85% accuracy on the USMLE, they still fail to provide accurate at-home diagnoses for the general public because of inadequate information and improper terminology. By introducing a toolkit designed to enhance patient interaction with AI, we can promote precise and structured inquiries to increase the reliability of LLM diagnoses. The toolkit optimizes user input through pre-determined prompt engineering, ensuring inquiries are precise and structured. By refining prompts for a professional tone and guiding users to provide additional relevant details, the toolkit enhances the quality of information provided to the AI. Simulated users from standard datasets were employed to test this approach showing this method significantly improved diagnostic accuracy compared to unstructured, casual prompts. Bridging the gap between AI and user interaction the toolkit aims to create a world where accurate self-diagnosis can make healthcare more accessible.
Maddy Brown
Project: Can I trust artificial intelligence with questions about my health?
Advisor: Qianwen Wang
Home Institution: Carleton College
Abstract: Large language models (LLMs) are becoming increasingly prevalent in society. They show especially promising potential in the medical field, where medical professionals and patients alike are using artificial intelligence (AI) platforms to fulfill or supplement their health-related inquiries. The performance of LLMs has been tested extensively on professional medical questions from the United States Medical Licensing Exam (USMLE) and other similar datasets. However, LLMs have not been tested on unprofessional medical questions that are likely to come from everyday patients turning to AI platforms for advice about their health. Our study aims to address this gap in the current research by investigating how the performance of GPT-3.5, a widely recognized LLM, on medical questions might be compromised if questions are asked unprofessionally or restricted to certain patient groups that are not covered in current datasets. By feeding GPT-3.5 baseline questions from the USMLE in addition to altered versions of the questions that are phrased in the tone of specific demographic communities, we calculate the difference in accuracy between the performance of GPT-3.5 on professional versus unprofessional medical questions.
Kaori Hirano
Project: Investigating the Relationship between Cybersickness and Motion Complexity in Virtual Reality
Advisor: Evan Suma Rosenberg
Home Institution: Carleton College
Abstract: This project studied the relationship between cybersickness and motion complexity in virtual reality environments. With between 20-95% of people experiencing cybersickness, it poses a significant barrier to the large-scale adoption of virtual reality and an equity issue to current uses. Motion complexity refers to how much a person’s overall movement can be explained by movement in different directions, such as up, down, left, or right. It was hypothesized that different levels of movement complexity would lead to different levels of cybersickness. The data analyzed were from a previous study that induced cybersickness in participants. A motion complexity measure, principal component analysis, and regression analysis were performed to determine the connection between cybersickness and motion complexity.
Shauntavia Hooper
Project: Embodiment and its Relation to Cybersickness Within Virtual Reality
Advisor: Evan Suma Rosenberg
Home Institution: Carleton College
Abstract: In the lab I was working in this summer, our overall focus was cybersickness within virtual reality. We wanted to analyze if a participant had the ability to see their virtual body by looking at a mirror if it would decrease their cybersickness. Due to this, I got to analyze the concept of embodiment and its connection to cybersickness. For further context, embodiment is how connected a person is to their virtual body while being inside said virtual body and controlling it within a virtual environment. I was able to create the virtual avatars that are used within the study, creating their movement, and adding virtual mirrors at checkpoints. Virtual reality is an ever-growing field that can greatly positively impact many different fields. For example, in the medical field, it can be used for training and therapy, and with the introduction of consumer virtual reality headsets, many people have begun entering virtual reality gaming. With the issue of cybersickness, many people who may benefit from virtual reality or could use it for leisure are essentially barred from doing so finding what may decrease cybersickness can aid in resolving this issue and make virtual reality more accessible to everyone.
Cynthia Maldonado
Project: Improving Object Detection On Autonomous Underwater Vehicles For Trash Detection
Advisor: Junaed Sattar
Home Institution: University Of North Carolina Charlotte
Abstract: Trash poses a great challenge in today's aquatic life and ecosystem. This project helps to safely remove trash without harming the environment by enhancing object detection in Autonomous Underwater Vehicles (AUVs). AUVs are self-powered robots used underwater for exploration and data collection. To improve AUVs we first have to create a dataset of underwater trash images from different aquatic environments. With the help of divers, we can utilize underwater robots equipped with cameras and sensors to capture trash that is commonly found in water. The images are labeled to identify each type of trash in a program called CVAT (Computer Vision Annotation Tool), an online image annotation tool. Hundreds of images are individually labeled using a bounding box to identify each trash item. Using YOLO (You Only Look Once), a state-of-the-art object detection system, we can develop accurate object detection models. By training these models on a large dataset, it can significantly improve the object detection capabilities of AUVs. This research seeks to become a long-term aid to help aquatic life and remove all types of aquatic debris in oceans, seas, and lakes.
Ashlyn McClendon
Project: Fusing Traditional Ecological Knowledge with Actively Studied Scientific Data in 3D Visualizations
Advisor: Daniel Keefe
Home Institution: American University
Abstract: To solve today’s most pressing challenges (e.g., climate change, sustainability) scientists and technologists will need to work hand-in-hand with the communities most impacted by these challenges. Our Indigenous research partners from Micronesia, many of whom have been forced to leave their islands and come to Minnesota, include master Micronesian navigators with historic and contemporary knowledge of fisheries, climate, and interconnected ecology, as well as scientists studying Pacific fisheries. In theory, these two ways of knowing the world (modern science, traditional ecological knowledge) should be perfect complements; yet linking quantitative computer-based scientific data with qualitative traditional knowledge passed on via ceremony, oral traditions, and experiential learning remains a major challenge. Our research aims to combine Indigenous Micronesian knowledge of the seas with Pacific maritime scientific data in a single 3D visualization that is both scientifically accurate and supports immersive first-person data experiences. Early results include implementations of two strategies for establishing common links, over space and/or time, to visualize relevant complementary data together. Results also suggest that mixing narrative and exploratory visualization and animated transitions may help users preserve contextual awareness in these complex visualizations.
Junior Mejia
Exploring Augmented Reality Aids for Cognitive and Spatial Assistance
Advisor: Zhu-Tian Chen
Home Institution: University of Central Florida
Abstract: This project investigates the potential application of augmented reality (AR) technology to support individuals with cognitive and spatial impairments, such as dementia. Leveraging AR's immersive capabilities and inspired by existing strategies to manage these challenges, the project aims to enhance daily living and task management. The application allows users to interact with spatial anchors and categorize them as "reminders," "items," or "locations" to achieve distinct functionalities. The project showcases how AR could potentially be tailored to create meaningful, real-world applications that address the specific needs of those with cognitive impairments.
Maura Putzer and Avani Papadopoulos
Project: The Impact of Individual Preferences on the Restorativeness of Virtual Reality Experiences
Advisor: Victoria Interrante
Home Institution: Smith College, Indiana University Bloomington
Abstract: Being immersed in nature has a beneficial impact on mental health. Virtual Reality (VR) can offer virtual access to naturalistic settings during events like dialysis and chemotherapy that prevent people from experiencing real natural environments. This study investigates the extent to which individual differences and preferences affect the perceived restorativeness of various environments, hypothesizing that high-biomass environments are universally restorative regardless of the individual. Our methodology involves three parts: (1) presenting static images of several natural environments to participants, (2) immersing participants in several VR natural environments, and (3) immersing participants in a single VR environment, either of their choosing or ours, following stressor tasks. Stressor tasks include mock job interviews and mental arithmetic. Following VR or image immersion, we will use surveys, such as the Perceived Restorativeness Scale (PRS), to assess the restorativeness of the environment. We will similarly use memory-related tasks, such as the n-back and backward digit span (BDS) test, in part 3 to determine the participant’s response to the VR environment.
Megan Williams and Devinn Chi
Project: NaloxSAVER: Overdose Detection System for Public Spaces
Advisor: Lana Yarosh
Home Institution: American University, Macalester College
Abstract: The opioid crisis remains a critical public health issue in the United States, with the Centers for Disease Control estimating over 105,000 overdose deaths in the past year. While significant efforts have been made in prevention and recovery, there is a pressing need for immediate, life-saving measures for individuals at risk of opioid overdose. Naloxone, a drug that can save lives if administered within a three-minute window, is essential in this effort. Our product, NaloxSAVER, addresses the gap in current resources by combining thermal infrared cameras with RGB cameras to detect opioid overdoses through monitoring breathing patterns. A live stream web interface includes an alert system when an overdose is detected, along with an informational video on how to administer the drug. We designed and implemented this novel system to provide rapid and accurate detection, achieving a 94% accuracy rate. We envision our product being implemented in public spaces where overdoses are frequent, such as metro and subway cars. By offering an immediate response mechanism, NaloxSAVER has the potential to save lives and improve outcomes in the fight against the opioid crisis.
