Sunggun Lee, a first-year student in Dr. George Pappas Lab, earned his bachelor’s degree in Biomedical Engineering from Duke University.
Lee’s educational background informed his prior research. “Most of the applications I had used AI in were for medical images such as MRIs or Calcium imaging videos for the brain,” Lee explained. This work entailed using AI and machine learning for medical imaging, specifically MRI scans at an internship at the National Heart, Lung, and Blood Institute (NHLBI) at the National Institutes of Health (NIH).
“It’s very complicated because there’s a lot of physics, math, and reconstruction that goes into it,” Lee noted. According to Lee, high magnetic fields and powerful magnets are employed to capture the images. A higher signal-to-noise ratio (SNR) generally leads to sharper, crisper images.
“Our idea was to go the opposite way,” Lee said. To do so, they would “use lower magnetic field strengths and apply machine learning techniques to reconstruct high-quality images.” Lee explained that this approach would allow the MRI to operate with weaker magnets that consume less energy, “usually MRIs have powerful superconducting magnets and complex cooling systems which make them heavy and difficult to transport.” The lighter, energy-efficient MRI would be portable and less expensive. “We could take it into underserved areas,” Lee said, adding scans in their current form remain very expensive for many people.
Coming to Penn shifted Lee’s focus from medical imaging to the physical world. “You always hear that robotics is the future of AI,” Lee said, adding that the recent boom in large language models (LLMs) has only accelerated interest in connecting AI to real-world systems. “Luckily, I got matched with Dr. Pappas,” Lee said, adding Dr. Pappas’ research interests extend to the intersection of physical robotics and AI. Part of Lee’s engagement with the research community at SEAS results from reading papers such as Jailbreaking Black Box Large Language Models in Twenty Queries The paper introduces Prompt Automatic Iterative Refinement (PAIR), an automated method for generating jailbreak prompts for large language models.
Jailbreaking, Lee admits “Is a bit of a scary term. “It basically means that you’re causing the LLMs to do things that they’re not supposed to be doing.” LLMs such as ChatGPT come equipped with safety mechanisms that prevent users from obtaining harmful or unethical content, like instructions for crafting a scam email. “Still, people sometimes find creative workarounds,” Lee said. For instance, in what is known as “role-playing”, someone might pose as a teacher demonstrating how to avoid scam messages. The “teacher” would then ask the model to generate a scam email as an example.
According to Lee, PAIR “uses two large language models in tandem. There is a target LLM,” he explained, “and an attacker LLM that iteratively refines its prompts to discover ways to jailbreak the target model. And they were very successful with what they did.”
“There’s a very collaborative atmosphere at Penn,” Sunggun said, adding that he finds both faculty and students supportive and encouraging.
When Sunggun’s not working, he enjoys both watching and playing soccer. “I want to join an intramurals team,” Sunggun said, adding, “I’m looking forward to next year’s World Cup. I’m glad that I’m here in Philly, where things are always happening.”