Advancing Scene Graph Generation with Open-Vocabulary Models

This research explores a major limitation in Scene Graph Generation (SGG)—the closed and predefined vocabulary that restricts its ability to generalize and recognize rare or novel objects and relationships. Traditional SGG models suffer from biases toward frequently encountered objects and poor performance on long-tail distributions, limiting their real-world applicability. To address this, I explored the integration of Large Vision-Language Models (LVLMs) with unrestricted vocabularies, leveraging zero-/few-shot learning and semantic generalization. This approach enables more flexible outputs, improved identification of rare or novel objects, and broader applications in tasks such as Visual Question Answering and robotics. My ongoing work focuses on solving issues with relation prediction and performing ablation studies to further enhance SGG performance.

Investigating Second-Chance Testing Regimens and Student Studying Habits

This research studies how different second-chance testing regimens affect student performance, learning, and course experience. We used student performance metrics from exams, quizzes, and homeworks from an engineering course taken by semesterly populations of second-year students to investigate this. In one semester, the course was administered with 90% of a student’s examination score being derived from their better attempt at an exam, and 10% of their examination score being derived from their worse attempt. In a different semester, the course used 2/3 of a student’s better score and 1/3 of their worse score to create the final score. Initial analysis shows a nonrobust significant difference between performance between testing regimens; my ongoing research aims to qualify this effect further through further analysis and comparison to student affective responses in each class.

Exploring the Effects of Frequent and Second-Chance Testing on Student Learning

This research investigates how different exam structures affect students’ learning and overall course experience. Conducted in a large computer science course, we compared three distinct testing regimens: frequent testing (11 exams with no second chances), infrequent testing (3 exams with second chances), and moderately frequent testing (4 exams with second chances). Our findings revealed no statistically significant differences in overall final exam performance, but frequent testing led to higher initial exam scores. Notably, students found frequent testing without second chances more beneficial for immediate learning but perceived moderate-frequency exams with second chances as less stressful and more supportive. This study highlights the nuanced trade-offs between frequent retrieval practice and remediation opportunities, suggesting that optimizing student experiences requires balancing exam frequency and opportunities for retakes.

Creating a Social Recommendation System for Connecting Researchers

The purpose of this research was to build a data-centered platform that aids researchers, particularly students and rising researchers, network with others at conferences (such as IEEE and ACM) based off their work and research interests. This platform aimed to increase inclusion and diversity among researchers and collaborators by emphasizing the works of women and non-binary researchers, among other underrepresented minorities.