Rahul Ramesh

Profile Image

I am a 5th year CIS Ph.D. student at the University of Pennsylvania advised by Pratik Chaudhari. I am interested in a principled data-centric view of deep learning. My research seeks to answer questions like: (1) What do “typical” learnable tasks look like? (2) How do we build optimal representations from unlabeled data? (3) How does data change over time and how should models adapt?

A theme the spans my work is the idea of task competition, i.e., all tasks need not share an optimal representation and we should instead train on data from a related subset of tasks.

I spent summer 2023 with NTT Research at the Harvard Center for Brain Science and worked on understanding why language models are able to solve many different tasks with Hidenori Tanaka. Previously, I interned at Amazon AI labs and worked on pre-trained models for image classification with Avinash Ravichandran and Aditya Deshpande.

I graduated from IIT Madras in 2019, at the top of my class, with a dual degree in computer science and a minor in physics. I was advised by Balaraman Ravindran and worked on building hierarchical representations for reinforcement learning using successor representations.

I am currently on the job market and am looking for research scientist and post-doc opportunities!



Select Publications

How Capable Can a Transformer Become? A Study on Synthetic, Interpretable Tasks
ICML 2024

A Picture of the Space of Typical Learnable Tasks
ICML 2023

Model Zoo: A Growing Brain that Learns Continually
Rahul Ramesh, Pratik Chaudhari
ICLR 2022

Deep Reference Priors: What is the best way to pretrain a model?
Yansong Gao*, Rahul Ramesh*, Pratik Chaudhari
ICML 2022

The Value of Out-of-Distribution Data
ICML 2023