Hi, I'm Dhruva. I research learning dynamics and inductive bias in neural networks. I'm currently a physics grad student at UC Berkeley. I'm a part of the Feature Lab. My advisor is Mike DeWeese.
news
01 Sep 26
I'll be a visiting researcher at Meta starting at the end of the month. I'm working with Andrey Gromov and others. We'll try to understand inference dynamics of autoresearch systems.
06 Jul 26
I'm at ICML in Seoul, sharing our work on the origins of Fourier-like feature manifolds in LLMs.
01 May 26
A bunch of researchers I respect a lot, and me, put out a perspective paper arguing that "There Will Be a Scientific Theory of Deep Learning," and it'll be physics-y in vibe. We refer to this approach as "learning mechanics."
04 Apr 26
I'll be at ICLR in Rio de Janeiro. We're sharing our work on tracing the inductive bias of kernel ridge regression all the way back to the raw statistics of the data.
28 Sep 25
I'll be a research intern at Google DeepMind this Fall. I'm working with Yasaman Bahri. We'll try to understand some aspects of residual stream geometry in LLMs.
research interests
At work, I'm mainly interested in the physics of (deep) learning. Here I'm using the term physics artistically rather than literally: in my search for analytical theories of deep learning, I value useful toy models, pragmatic approximations, illuminating limits, and convincing quantitative match with practical experiments.
I'm especially curious about feature learning in language models. Language has (surprisingly?) nice structure, and there's mounting evidence that sequence models encode some of this structure in simple and understandable ways. I want to understand how gradient-based training cultivates "circuits" in the weights from structure in the data. (One can give an operational definition of a circuit via its inference time behavior).
I'm disheartened by the proliferation of AI slop/misinformation and worried about the threat of human disempowerment. I oppose the concentration of capital and lobbying power at AI companies. At the same time, I'm optimistic about the potential for AI to accelerate science, improve medical outcomes, and teach things to curious people. I hope that my research helps to enable the latter without contributing to the former.
teaching
I really enjoy teaching — I think sharing knowledge is dope. Early in my PhD I won the Outstanding Graduate Student Instructor Award from the Physics department, as well as the selective Teaching Effectiveness Award from the university. You can read my award essay here.
Here's some anonymous student feedback I'm extremely proud of:
- Dhruva is really chill and relaxed and an easy person to talk to. He is very prepared for discussions and they are really helpful in learning the material. He seems to be good at physics and is always willing to help/explain in the best way that he can. He also tries to explain the basics of a concept at the start of a discussion, which will help a lot because some concepts are very murky during lecture. Always learned a lot after attending.
- Very approachable and open to being asked questions, has good mini–lectures, I found his jokes funny
- Dhruva is an excellent TA. He starts each discussion with a small presentation about topics covered in the professor's lecture and/or topics he believes important to cover. These are easily understandable and often are more useful than the lecture due to Dhruva's excellent teaching style. He listens to the students and encourages their participation as well as questions. He relates to students as people and always comes prepared, even when faced with family emergencies. preparation and organization, attitude toward students, responsiveness and accessibility
- Dhruva is very prepared and understands what students' are struggling on. He understands the subject well and is good at teaching.
- Dhruva is an incredible teacher, and sparked an interest in physics beyond the scope of this class everyday. He put our equations in perspective.
- Dhruva is very knowledgeable, creates his own questions that are interesting (questions could be made a little harder), and very effective in explaining the solutions. He comes to class prepared by having spent time creating and solving questions beforehand. Due to this preparedness, Dhruva is a very effective teacher. Dhruva is also receptive to feedback and enthusiastic.
- Dhruva is very organized. He conveniently keeps all the notes and discussion problems on one website. His notes are clear and concise and if/when he doesn't have the answer to a discussion problem, he can still solve and teach it on the spot. He changes the difficulty of the problems according to the needs of the students.
- Going to discussions with him was really helpful. As someone with hardly any physics background, he made the material easier to understand and would consistently be engaging. I honestly do not think I could have passed the class without him.
- I was a student of his, and he was able to present material excellently in conjunction with lecture to ensure that we as students was able to fully comprehend material taught and ensure our success on exams and test questions. He was really able to teach us through a student perspective. His discussion and lab sections were always insightful with creative and fun questions or jokes to lighten the mood on the often times dull topics. I would not skip his section!
I was a Graduate Student Instructor for Physics 7A in Fall '21, Spring '22, Summer '23, and Summer '24. I've also taught Physics C10 and Physics C21 several times. I was also the graduate representative of the Physics department's Major Curriculum Committee.
random facts
Some activities I enjoy:
- cooking for friends
- playing vibe chess
- messing around with software synthesizers (any Arturia Pigments fans in here?)
- observing plants, tending for plants, watercoloring still life (plants), going on long chill walks (to witness some plants, of course)
- doing flipturns in a pool (the swimming part in between is alright too I guess)
Some bands and tv I enjoy:
- Parcels
- Khruangbin
- Men I Trust
- Jungle
- Dune 2021
- Princess Mononoke
- Arrival
- Andor
Before moving to Berkeley, I studied physics, computer science, and astronomy at UT Austin (hook 'em).