Oishi Banerjee

Year of Graduation: 
Oishi Banerjee

“One of the highlights of my Exeter career was a meditation I presented at Phillips Church about how Exeter allowed me to pursue different academic passions.”

Artificial Intelligencer 

If there’s ever a Latin-speaking robot, there’s a good chance Oishi Banerjee will be the brains behind it. Banerjee, a “deep learning engineer” who helps design household robots with 3D vision at Matician, earned a Classical Diploma at Exeter and first came to appreciate computer science in a required introductory course with Instructor Brian Sea. That robot will likely sing, too: Banerjee studied opera at the Academy and won the 2013 Vocal Competition performing “Una donna a quindici anni” from Mozart’s Cosi fan tutte. “One of the highlights of my Exeter career was a meditation I presented at Phillips Church about how Exeter allowed me to pursue different academic passions,” she says.

At Stanford, Banerjee majored in Greek and Latin literature — a minor in computer science introduced her to the world of artificial intelligence and natural language processing — and earned a master’s in computer science. Looking to do more applied research, Banerjee found a position with a Stanford lab using AI to answer medical questions. Her humanities background was appreciated by research colleagues, who invited her to co-produce “The AI Health Podcast,” a podcast exploring how AI transforms health care and biotech. “I’m always doing 20 different things at the same time, so opportunities just pop up,” she says.

We caught up with Banerjee at her home in the Bay Area to hear about what’s next.

What’s on the horizon for AI in medicine?

AI has the potential to allow scientific discoveries at paces we’ve never before imagined. As an example, AI can design drugs that target antibiotic-resistant pathogens, are less toxic, have fewer side effects, and are more likely to work, so less money will be wasted on expensive trials that often fail. Within the next 20 years, AI has the potential to revolutionize how we do drug development and cut medicine costs. Another example is using the same techniques that power devices such as Alexa to help doctors with electronic health records, freeing them to have more time with their patients.

Are there any downsides to AI?

In order to make a modern AI system, you give it data that already exists and you have it learn patterns. It’s hard to understand what patterns AI has learned, so there’s the risk of learning bad patterns, even picking up on racial or gender bias in the data. In some dermatology studies, a lot of the training data points were taken from lighter-skinned patients, and you can see a decrease in performance when the model is deployed on pictures of darker-skinned patients. Even if researchers aren’t actively being malicious along the way, it’s so easy for these AI models to learn patterns that are prejudiced somehow.

What are your long-term goals?

At Exeter, I loved being a Latin tutor and I loved being at the Harkness table. … I taught English and computer science at Stanford. I think I’m going to be teaching regardless of my official job title. I’m passionate about making somewhat opaque academic material easier to understand. I want to spread the knowledge. 

— Sarah Zobel

Editor's note: This article first appeared in the winter 2022 issue of The Exeter Bulletin.