The Summer Undergraduate Research Program (SURP) enables students to conduct extended, focused research in close cooperation with a Pomona faculty member. Below are recent summer research projects in the Computer Science Department.
Automatic Generation of Health-related Questions
Vivien Song ’25, Advisor: David Kauchak
The summer after sophomore year, I worked with Professor Kauchak to research new algorithms for automatically generating health-related corpora containing multiple-choice questions. We wanted to generate corpora with known and fine-grained levels of difficulty and understandability: to do so, we compared two approaches for generating MCQ questions and wrong answer choices, using a collection of automated metrics to understand the difference between those approaches and later validating our results through user studies. When I first began researching, I was new to natural language processing and reviewed existing literature to understand current state-of-the-art techniques. For the rest of the summer, I evaluated ChatGPT-generated question-answer sets against search-engine-generated sets through n-gram similarity based metrics (BLEU, ROUGE-L, coverage, unigram precision); a novel question taxonomy I created, based on textual diversity and question type, to address overall difficulty; and human annotations of question-answer set quality, finding that there were meaningful ways in which automatic ChatGPT-generated sets supplemented existing search engine techniques. I ultimately had the opportunity to present my work at the 2024 AMIA (American Medical Informatics Association) Symposium; and later published alongside Professor Kauchak, his collaborators, and two other Pomona research students at the 2024 IntelliSys (Intelligent Systems) Conference.
As a freshman, I wouldn't have thought that I'd get such an opportunity, and I wholeheartedly believe that such experiences like this are what make Pomona CS so unique. Professor Kauchak's guidance and support empowered me to explore different research interests I came in with, bridging my backgrounds in English literature and computer science. The individualized mentorship I received over the summer pushed me to ask questions with ease and eagerness, approach seemingly unapproachable topics, and find new ways of contributing wherever I am. I've learned an incredible amount from just one summer and I can't wait to return to research in the future.
Text Simplification for Medical Texts
Robin Shen ’25, Advisor: David Kauchak
Over the summer of 2024, I worked with Professor Kauchak (also known as Dr. Dave) to investigate text simplification and its applications towards improving the accessibility of medical text to a general audience. Specifically, we wanted to create an efficient method to evaluate how well current large language models such as ChatGPT were performing at this task, both in simplicity and similarity - two key components of a good text simplification - as well as if there were factors that made for better or worse simplifications.
Over the course of the 10 weeks, we tried many different approaches and ideas. For me, the most satisfying idea and outcome built on a class I'd taken with Dr. Dave the previous semester. I designed a dynamic programming algorithm to efficiently align sentence pairs between an original text and its simplified version. First, we did some manual feasibility testing to ensure that a dynamic programming algorithm would be suitable. Following this, we coded the algorithm in Python - which evaluates the best pairings of sentences at each local position based on similarity scores from Google's BERT model. The algorithm was fairly accurate, with almost perfect precision and recall when compared with the pairings derived from manual testing.
We also evaluated how much relevant information loss was happening in our simplifications by seeing how many key concepts were being removed in the final simplification. We found that using the sentence alignment algorithm, sentence pairs with a similarity score lower than 0.94 were much more likely to have lost important relevant information - showing promise in the algorithm for helping to identify overly aggressively simplified sentences. Eventually, one interesting application of this research is that this could be used as a tool to help writers in simplifying text by automatically flagging overly simplified sentences.
Sim-to-Real Transferability
Christy Marchese ’24; Advisor: Anthony Clark
For the past 2.5 years, I’ve worked with Professor Clark in the Autonomous Robotics and Complex Systems (ARCS) Lab, researching methods to bridge the sim-to-real transferability gap for mobile robots.
I first began researching with the ARCS lab the summer after my freshman year at Pomona. That summer, the group and I explored various neural network architectures and data collection techniques for autonomous navigation in simulated environments, looking at what architectures and methods we could utilize to best enable a robotic agent to learn the task of navigation through a simulated maze. For that project, I developed novel neural network architectures that explored varying uses of state (memory), one of my architectures being a hybrid convolutional neural network that took inputs of an image and categorical data (the previous output) to inform its navigational decision making. I was also able to take on a variety of tasks in the lab from writing python automation scripts for the training of all our different models to soldering wires for the electronics of our mobile robots to designing robot parts with CAD. We ultimately published our work from that summer, titled Investigating Neural Network Architectures for Navigation in Simulated Environments, at the IEEE’s SSCI conference in 2021.
These days in the lab, I have taken a much more active role in driving new research questions, proposing new methodology to approach these questions, and exploring more novel ideas of sim-to-real transferability. I am also driving the exploration of adversarial robustness in the lab, looking to bridge my research interests and experiences in security and the security of machine learning to approach improving model robustness and sim-to-real transferability in novel ways. I feel incredibly blessed and grateful that the lab has supported and empowered me with the freedom to explore my many different research interests, projects and opportunities. My experiences in the lab have really helped me find a passion for research and the approach of difficult questions. So much so that now I intend to pursue my interests of security and the security of ML in graduate school through a Ph.D. in computer science, and I move onto this next part of my research journey with hope and excitement for the future.
Animating Asterism
Katiana Wieser ’24; Advisor: Joseph Osborn
Starting my freshman year, I have worked with Professor Osborn on Asterism, exploring the idea of how to create a game engine engine and lowering the barrier to entry when designing video games. Asterism is a library implemented in the Rust programming language and has been used to implement two different game engines. When I first began researching, I focused on orienting myself with the existing work and Asterism library because everything was completely new to me–the idea of Operational Logics (the basic features used to create games), the Rust programming language, and what it looks like to conduct computer science research! However, over time and with the help of Professor Osborn and a more senior researcher, who always took the time to answer my many questions, I found my footing.
At first, I worked more on understanding and exploring the flexibility and uses of game engines. Then I transitioned to demonstrating how some of the theoretical ideas I had explored could be implemented through the existing Asterism library by making games in a game engine made by previous student researchers, which brings me to my personal pride, which was adding more advanced animation capability to the library. It was my interest in animation and the intersection of art and computer science that led me to Professor Osborn’s lab. I was excited when the opportunity arose to combine an area of personal study with my research. Most of the games in Asterism used simple graphics, where all visuals were single-colored geometric shapes; my contribution was to implement the basis of a sprite-based animation system. Using my animation components, I was able to add flip-book style sprite sheet animations in two of the games I had previously made in the engine. I enjoyed having the opportunity to do a little digital drawing when making the sprites.
The lab’s work on Asterism resulted in “Asterism: Operational Logics as a Game Engine Engine,” a paper that I coauthored alongside Professor Osborn and another Pomona student researcher. We presented the Programming Languages in Entertainment (PLIE) 2021 workshop at Artificial Intelligence and Interactive Digital Entertainment (AIIDE). In my time conducting research at Pomona with Professor Osborn, I have learned so much and know I have also only begun to scratch the surface of what is possible. My experiences helped and empowered me: seeing how I can contribute, giving me the confidence to speak up more and not be afraid to share my ideas and perspectives in all areas of my life.