Achyuta Rajaram won this year’s Regeneron Science Talent Search $250,000 top prize for his work on making machine learning more efficient and safer.
Every year, the Society for Science hands out over $1 million to the competition’s finalists. This year’s winners were announced at a gala on Tuesday, and Rajaram was so surprised he won that a security guard became concerned.
“He was worried that I’d just faint there and then,” Rajaram told Business Insider.
The 17-year old Phillips Exeter Academy student created a project around improving the speed and accuracy of computer vision models, which he told Business Insider can function as mysteriously as ChatGPT.
“If you ask it a question and it gives you an answer, you don’t actually know what happens in between these steps,” he said of the chatbot adding that, “problems can lie lurking within these models” that eventually go on to make mistakes.
Understanding more about how an algorithm gets from input to output will not only make it faster but safer, Rajaram said. And since machine learning is the foundation for AI, Rajaram’s project could make AI smarter and faster, too.
Building faster, safer AI
When humans look at an image, a cluster of neurons in your brain will light up in response, Rajaram said. “The same thing happens in these computer vision models.”
So how does a computer tell the difference between a car and a cat? If the model is looking at a cat, it might begin by focusing on the ears, nose, and whiskers — key features that distinguish it from a car.
This thought process can be broken down into, what are called, circuits — the part of a model responsible for detecting different parameters.
Typically, if you want to find a specific circuit in a computer’s brain, you have to do it manually, Rajaram said.
“But this is extremely, extremely impractical, especially as models get bigger to the billions of parameters,” he added. So, he built an algorithm to automate the process, for a smarter, faster model.
One part of Rajaram’s project involved looking at an open-source model that can read text from images. That ability makes it vulnerable to a “very, very strange adversarial attack,” Rajaram said.
If an image of a red traffic light is next to text that says green, the model would often classify it as a green traffic light. “Models doing these misclassification errors is certainly problematic,” he said.
Rajaram was able to isolate the model’s text reading capability and remove it, “repairing this attack and protecting” it against that specific failure, he said.
Beating out 2,000 competitors
Participation in Regeneron Science Talent Search has grown and shrunk over the years, reaching its peak in the late 1960s during the Apollo missions. This year had the highest number of applicants since that decade, Maya Ajmera, the Society’s president and CEO, told Business Insider.
Scientists and engineers evaluated over 2,000 applications and narrowed them down to the top 300. “These 300 scholars are the top scientists and engineers,” Ajmera said. A panel of about 20 then selected the top 40 finalists.
Unlike with a science fair, STS judges focused on candidates’ achievements beyond their projects. They asked students with physics projects biology questions, for example. The purpose is to test their knowledge in other fields.
“They want to see how well-rounded you are because we are looking for the next generation of scientific leaders,” Ajmera said.
Over the last 82 years, STS’s alumni have included 13 Nobel Prize winners, 23 McArthur Fellows, and 11 National Medal of Science awardees.
“I can sleep better at night knowing that these young people are trying to solve the world’s most challenging problems,” Ajmera said.
Rajaram plans to continue studying computer science at MIT in the fall. His advice to anyone who wants to apply for the Regeneron Science Talent Search is to “be really, really curious about everything.”
That includes subjects outside your chosen field, he said. “I am a personal believer that almost everything is infinitely cool if you go deep enough into it.”
Read the full article here