Rosie Zhao

Hi there! I’m a third year PhD student in Computer Science at Harvard University advised by Boaz Barak and Sham Kakade. My research is largely focused on the scientific study of deep learning. I’m grateful to be supported by a Kempner Institute Graduate Fellowship.

I previously completed my undergrad and Master’s at McGill University under the supervision of Prakash Panangaden; there, my research was geared towards representation learning and state abstractions in reinforcement learning, as well as various topics in theoretical computer science.

My socials can be found below.

news

Jul 21, 2024 I will be attending ICML 2024 in Vienna! Come check out our spotlight paper about the implicit bias of SGD noise in online learning on Wednesday, as well as the High Dimensional Learning Dynamics Workshop on July 26! :airplane:
Jul 10, 2024 New preprint on evaluating optimizers for LLM training across different hyperparameters and model scale! See the paper, blog post, or tweet thread for our ablations and empirical investigation on the importance of preconditioning for adaptive optimizers like Adam. Done in collaboration with Depen Morwani, David Brandfonbrener, Nikhil Vyas, and Sham Kakade. :zap:
Jun 8, 2024 Our paper “Policy Gradient Methods in the Presence of Symmetries and State Abstractions” is officially published in JMLR! This paper is the culmination of work done during my Master’s on leveraging MDP homomorphisms to make state-action abstractions in continuous spaces. Done in collaboration with Sahand Rezaei-Shoshtari, Prakash Panangaden, Doina Precup, and David Meger. :sparkles: