Rhodes Scholar digs into deep learning

After returning to Cambridge from a work trip to New York City, then-sophomore Neil Band walked into his dorm room and found his roommate in the throes of a terrifying epileptic seizure.

“Not only was he having a seizure, he was actually choking on something. I had to give him the Heimlich maneuver,” said Band, A.B. ’20, a computer science concentrator at the Harvard John A. Paulson School of Engineering and Applied Sciences and a 2020 Rhodes Scholar. “If my train from New York had been delayed, I don’t know what would have happened. That was super-jarring, and I was absolutely shaken by it.”

That experience inspired Band, who was completing a summer internship at data analytics firm Kensho Technologies, to explore whether machine learning could be applied to seizure prevention.

While that task had already been addressed, Band found that the commercial devices to help predict and prevent seizures have a number of pitfalls. The most popular device, a bracelet that measures electrical impulses, hasn’t caught on because it is notorious for overheating and burning people’s skin, Band said.

“Ultimately, that problem arises because it is just too computationally expensive to run that algorithm on that device right now,” he said. “I started to wonder what could make that model take up less energy and compute time. That was a breakthrough moment for me because I started to realize how tangible deep learning could be.”

Band had harbored an interest in computer science since his childhood in Omaha, Nebraska. After taking an AP computer science course in high school, he was hooked. As a freshman, he enrolled in Introduction to Computer Science (CS50) and soon saw how the discipline could be applied to solve real-world problems.

Band took his CS50 project, an app to help students find impromptu events on campus, through incubation at the Harvard Innovation Labs. While the startup didn’t go anywhere, it inspired him to use computer science to make a difference in the world.

During his summer internship at Kensho, he delved deeper into algorithm development, working on a natural language processing technique that used massive amounts of scraped data to get a very domain-specific understanding of words in a corpus of text.

After the seizure episode, Band doubled-down and reached out to his concentration advisor, Associate Professor Stratos Idreos, looking to take as many systems and databases courses as he could, including Idreos’ course Data Systems (CS 165).

“That course was definitely a turning point for me. I had always thought I would do a startup or work in industry,” he said. “That was the point at which I started to understand that academia is really fascinating and can have this huge impact on the world.”

Band began working on research projects alongside Idreos, digging deep into neural networks in an effort to understand how they really work. He worked on novel strategies to develop massive algorithms that could be optimized to minimize both computation time and memory usage.

The work hearkened back to his original interest in developing a less-computationally expensive algorithm for seizure detection devices. But he’s also seen how he can carry it forward and apply the same principles to new areas, like drug discovery, which he wants to focus on at Oxford University next year.

“It currently costs about $2.6 billion and takes an average of 12 years to develop a new drug. If you have a terminal condition or a devastating illness, it is often not feasible that the drug will even get developed in your lifetime,” he said. “These big pharmaceutical companies have to spend a lot of time physically testing different versions of a chemical they think will work. A lot of it is just trial and error.”

Machine learning offers great promise to streamline that process, Band explained. An algorithm could run through thousands or millions of chemical combinations in a fraction of the time, targeting the best manner in which certain small molecules could be combined to inhibit a particular disease target.

Band isn’t intimidated by the unparalleled depth and breadth of a problem that has plagued researchers for decades. Rather, he sees it as a golden opportunity to apply his skills in an area where he can make a real difference.

“There has never been a drug that was produced computationally—that was procedurally located by an algorithm—which has ever helped a patient. This is a green field,” he said. “I think it would be fascinating to work on algorithms to speed up parts of the drug discovery process, eventually towards the goal of producing something that could be much cheaper, and have tangible, positive impacts on many people.”