I am a first-year Ph.D. student at Stanford NLP advised by Dan Jurafsky and Christopher Potts, and currently rotating with Noah D. Goodman. My research is focused on interpretability.

I want to understand how neural networks (particularly language models) work. I believe that this is a tractable goal that can be accomplished in my lifetime. Some things I’ve been thinking about recently:

  • How can we ensure that explanations of model behaviour are actually faithful? New methods grounded in causal inferences are promising, but we still need more theory, benchmarks[AJP ‘24], metrics, etc.
  • In a self-supervised learning world, can linguistics still be useful in guiding how we do interpretability on language models?[AJP ‘24]
  • Can interpretability provide actionable findings that help us make better models? Control over model behaviour[WA+ ‘24], greater robustness, etc.

Machine learning is still a kind of alchemy. We should turn it into a science. To that end, I am inspired by work in NLP, causal inference, information theory, and psycholinguistics.

Besides doing research, I enjoy eating spicy food, (attempting to) play basketball and climb, and listening to $\{$Indian music, rap$\}$. And if you handed me a violin, I would probably be able to make some sounds.

If you want to chat about research or life, feel free to send me an email (aryamana [at] stanford [dot] edu)!

Brief history

I was born in New Delhi, India, raised in Savannah, Georgia (the U.S. state), and I think of home as Washington, D.C.—where I spent part of high school and my undergrad. I’ve wanted to move to the Bay Area for a long time, and I’m glad I made it here!

Before coming to Stanford to start my Ph.D. in 2023, I completed my B.S. in Computer Science and Linguistics at Georgetown University. There, I was mentored by Nathan Schneider as a member of his research group NERT. In those days, I primarily worked on computational linguistics and did a lot of linguistic annotation for Indian languages. Regardless of what I currently work on, my research style is probably largely copied from Nathan’s.

Since 2021, I have also been closely working with Ryan Cotterell at ETH Zürich on information theory, and I visited Switzerland in Summer 2021 and 2023. From working with Ryan, I have learned to be a little less scared of doing math.

In 2022, I spent the summer at Apple in Seattle with Robert Daland working on evaluating robustness on a ton of languages for Siri, and winter at Redwood Research in Berkeley working on mechanistic interpretability.

In light of sudden advances in AI in late 2022, my research interests pivoted significantly towards interpretability, but I still have a deep love for language(s).