Achyuta Rajaram, Alan Bu, and Riya Tyagi Win Awards at Regeneron Science Talent Search

By ARYAN AGARWAL, MARVIN SHIM, KEVIN THANT, and ANDREW YANG

The Regeneron Science Talent Search is one of the most prestigious STEM fairs for high school seniors in the United States. Thousands of students annually submit their research projects in various fields of science and mathematics. Three hundred semifinalists are then selected, based on the quality and originality of their research, to present their projects to a panel of judges. Only forty national finalists are chosen to attend a week-long event in Washington, D.C., where they present their research to leaders in their respective fields, gaining incredible experience and advice from the brightest minds in the country. This year, three Exonians were selected to become finalists at STS, marking an incredible achievement for each one of them. Seniors Achyuta Rajaram won the first prize of $250,000, Alan Bu won the tenth prize of $40,000, and Riya Tyagi won $25,000 as a finalist. Each of their projects was incredibly intricate, the product of hours of hard work and dedication.

Rajaram was the winner of the entire competition and his project contributed to a widely discussed field in today’s world: the interpretability of machine learning models. “Machine learning models have seen wide adoption in vision-based systems from tumor diagnosis to driverless cars,” he explained. “However, the increasing reliance on such models for automated decision-making in the real world has raised questions about their comprehensibility to humans. What do vision models actually ‘see’ is the question.” 

By deconstructing the “circuits” within these large models, Rajaram believes that humans could learn more about how machine learning is able to give us incredible results, despite us not knowing how. He tested his theory with a vision model by attempting to remove “the spurious correlation between images of the text ‘green,’ and the color of traffic lights.” This was challenging work, as this specific method had not been tested, so there was not much previous research to build off of. 

Rajaram in fact describes the most difficult part of his journey as “figuring out how to evaluate the method. With no existing ‘correct’ examples of circuits, most of the work was about quantitatively understanding how ‘good’ the automated circuits were.” 

Many might wonder what prompted him to take on such a project, and how he managed it. “I have always been interested in computer science, and the nature of intelligence in general,” Rajaram said. “I have a hunch that the most impactful field towards the end goal of ‘reverse engineering intelligence’ is actually interpretability. I felt that neural networks are the natural place to study the algorithms behind intelligence; we can do ‘surgery’ on neural networks; with direct edits, we can perform intervention studies which would be impossible on humans. I believe that this specific project, of automatically discovering circuits, ‘follows naturally’ from other circuits-style work in the field, like Arthur Conmy’s work in language models.”

He additionally found great mentorship in Dr. Sarah Schwettmann from MIT CSAIL. “She was extremely instrumental to my growth,” said Rajaram, “both throughout the project, and supporting me during the finals week process itself. One moment that still sticks with me was the letter she provided, telling me to ‘go kill it at STS.’” 

Additionally, he commented on his experience with Exeter academics: “Most of the work was completed over the summer before my senior year, so I didn’t have to do much balancing. In general, I believe that with the academic rigor of Exeter, it’s important to prioritize and focus on doing what you find most interesting.”

Rajaram reflected on his experience of winning, and what he would like to do in the future, saying, “I was mostly just shocked! All of the finalists were extremely talented; I wouldn’t rank myself in the top ten, much less first. I think one of the security guards thought I was going to faint on stage, which was pretty amusing.”

  But this is not the end of the road for Achyuta’s research. “I plan to continue working on automated interpretability research at Torralba lab. I hope that this work becomes a “tool in the toolbox” of interpretability research.”

Bu’s project researched spanning-trees within a graph, a mathematical concept, but one which has several practical applications in fields such as physics. When describing his project, Bu explained it as putting yourself in the shoes of a railroad conductor, drawing out a map of several cities which represented vertices in his graph, as well as edges which represented “connections” or railroads. “Let’s say you’re also a very greedy railway conductor, so you don’t wanna build any extra roads that aren’t necessary,” said Bu. “It turns out we can abstract this idea in mathematics where this entire system is called a graph.” He called a minimum spanning tree a version of the graph where one can travel from any “city” to another using the least amount of edges needed. 

In particular, Bu studied how “the number of different ways you can build this number of roads signifies how complex the graph is. If there were a billion different ways to build this road system, it’d be really complicated. But if there is only one way, it’s not very complicated.” 

Bu addressed how this idea was originally posed by physicists who were “modeling ferromagnetic systems as atomic lattices where the number of spanning trees in those lattices represent certain physical properties of the lattice like the energy or the entropy and so on.” According to him, the question of a graph’s complexity is incredibly difficult to tackle, which is why his research is so astonishing: he connected “two problems which don’t resemble each other at all,” to allow him to approach his questions. Matrix determinants, another part of mathematics found in linear algebra, showed great promise to Bu, where much of his project was based. 

The most challenging part of this project was that nobody has proven the full conjecture yet. He was able to discover a one to one correspondence between connections within a planar graph and the determinant of a matrix by months of dedication. 

Bu described, “As research always is, there are lots of moments of inspiration. Each time you think there’s nothing more here, and you sit for a while, you then realize that there’s something more. There’s always something going on at a deeper level that you haven’t discovered yet. And every time you discover one piece of information, you will be looking for another. It’s never just given to you, right? There’s no paper out there that says: this equals to this and you should explore it.”

The research was only part of the challenge involved with STS. Being selected as one of the top forty participants and exhibiting his work in Washington during the spring break proved to be just as demanding. 

According to Bu, he had to give speeches one after another during the public period, thinking that he wasn’t going to win when he had comparatively less judges come to him during the judging session. “There were crowds of people,” he recalled. “It was crazy. There was a hoard of people. We were there for two hours, presenting our topics. By the end, none of us could speak because our throats were parched from talking too much.”

Bu continued, “On public day, they have judges that visit your poster to see how good your presentation is. At the very beginning, I had very few judges, the least out of everyone. But then they had a lunch break, and suddenly there was a surge of judges coming to my poster. I didn’t know what that meant because everyone else had lots of judges. I was like, it’s probably over, but I’ll keep doing my best because you do your best and in the end, whatever happens, happens.”

In the end, Bu placed tenth in the competition. He reflected that “In research, there’s plenty of times that you think you’ve proven something, but actually you haven’t. In my research project, two and a half weeks in, I actually thought I proved the entire thing. But then I discussed it with my mentor and as I was explaining it to him, I was like, oh my god, this doesn’t work. So it was a huge setback. But when things like that happen, you just stick through, and you give it your all. Even if it doesn’t work in the end, the research itself was a fun enough experience. I feel like if you haven’t given it your all, then maybe in the end you will regret not giving it your all. But if you have given it your all, at the end, you really can’t complain about it. I think it’s the best way to go about things.”

Finally, Tyagi conducted a project in “investigating the mechanism of algorithmic bias in medical image settings. It’s actually pretty similar to Achyuta’s project,” she said, “just from a different angle. The whole goal was, like all AI interpretability projects, to understand why AI models are making the decisions they’re making.” More specifically, she “was responding to a study in 2021 that found that AI can learn a patient’s race and ethnicity from medical images that seem to contain no such indications of race or ethnicity.”

Tyagi shared that she gained her inspiration from reading an article about algorithmic bias in the ninth grade. “It almost felt terrifying to think that this idea that I’d idolized so much was actually causing harm to so many people at this moment,” she recounted. “People who looked like me as well. And one thing that was one of the factors, too, because at least with my project, what I actually found was that AI models perform more poorly. You can fix these AI models, but this particular AI model I was looking at performed more poorly on patients of color because of differences in darker skin pigmentation, which is a darker cordial pigmentation, which is pigmentation of the choroid, which is part of the eye behind the retina, which appears in retinal scans.”

Tyagi confessed that the most challenging part of her journey was “trying to figure out and trying to explain what was going on. And the reason is because there was actually previous research that had been done, trying to understand why AI could learn race or ethnicity from these images. And they actually found nothing. So they looked at tons of biological features, like bone density or body mass index or BMI or things like that, but none of them had any correlation whatsoever. The end result of the paper was them saying that they had no idea what proxies in these images are responsible for AI’s ability. And so for my project, I wanted to try to find another way in.”

“So rather than considering which elements of the biology could be related to race and could be confounders, I looked instead at which image features might be connected to race somehow. And that’s how I ended up discovering that rather than biological confounders, there was actually an AI model that was performing more poorly on people of particular races.”

As to Tyagi’s determination, she described how she balanced Exeter academics with her STS project. “I think one thing that’s really cool about STS and a ton of other science fairs as well, is you don’t need to spend too long on your project.” 

She continued, “So I usually worked full-time during breaks and also during summer. I did a big initial part of this work over summer break in tenth grade. Then I actually had all the code for it and I ended up making tons of modifications, writing the code, etc. A little bit in eleventh grade, in the summer of eleventh grade, and that’s also when I wrapped up my paper. To be honest, I didn’t spend too much time during the school year.” 

Tyagi finally reflected on a keynote speaker at STS who “said one thing he had really focused on was pushing beyond the norm, pushing the boundaries of what people thought was possible. And not in the cliché way that we always talk about, but rather almost this idea that if you’re failing, that just means if you’re not getting news coverage, if your project or if your research or whatever work you’re doing is not gaining traction, maybe the world isn’t ready for it yet.”

Ideas. Some ideas may never take fruition because some people did not have the means to do so. But only hard work can make ideas become reality, and that is what STS allowed these three Exeter students to do. In the end, the many hours they dedicated to their impressive projects paid off, not through the awards but by knowing the fact that their work could seriously impact the world. These three incredibly talented Exonians exemplify the benefit of working hard, persevering through adversity, and seeking to push the boundaries of knowledge. That, in the end, will be what sends our world towards the better future that we envision for humanity.