Jeffrey J. Ma

Science and Engineering Complex (SEC) 5.101

150 Western Ave

Boston, MA 02134

About Me

Welcome to my site! I’m Jeffrey Ma: I’m currently a Harvard CS PhD student, advised by Prof. Vijay Janapa Reddi. I’m interested in the intersection of machine learning, systems, and multi-agent interaction. I am especially interested in the following:

Efficient model, data representation, and learning.
Continual learning and scalable methods of skill acquisition in large foundation models.
Incentive-aligned systems of multiple learning agents.
Parallelism, asynchronicity, and resiliency in ML systems.

I’m currently working on LLMs for code, studying how we can improve correctness in LLM code generation, incorporate compiler fundamentals, and optimize code scalably for both asymptotic runtime and hardware specific performance in collaboration with Dr. Amir Yazdanbakhsh.

I previously attended the California Institute of Technology for my undergrad studying computer science and finance, where I was advised by Prof. Adam Wierman and worked with Prof. Animashree Anankumar, Prof. Yuanyuan Shi and Prof. Florian Schäfer on competitive optimization methods in multi-agent reinforcement learning settings.

I’ve also worked in industry, both as an intern and full-time prior to my PhD:

At Google Brain, on the TensorFlow Extended team, working on MLOps to continuously train models on newly arriving data.
At a self-driving startup, Nuro, on the ML Infrastructure team, on post-training model optimization and deployment.
At a quantitative finance firm, Citadel, on the E-Trading and Order Management System (OMS) teams, working on algorithmic and automated methods of trading and booking fixed income instruments.

news

Apr 10, 2024	I’ll be interning at Amazon AWS AI Labs in NYC this summer working on fault resiliency in LLM training! Definitely reach out if you’re in the area and want to chat!
Aug 28, 2023	Started as a PhD student at Harvard, working on ML + systems, large language models, and code generation!

latest posts

May 03, 2024	(WIP) Learning to Learn (Efficient ML Seminar Series)

selected publications

Polymatrix Competitive Gradient Descent

Jeffrey Ma, Alistair Letcher , Florian Schäfer , and 2 more authors

Nov 2021

Abs arXiv

Many economic games and machine learning approaches can be cast as competitive optimization problems where multiple agents are minimizing their respective objective function, which depends on all agents’ actions. While gradient descent is a reliable basic workhorse for single-agent optimization, it often leads to oscillation in competitive optimization. In this work we propose polymatrix competitive gradient descent (PCGD) as a method for solving general sum competitive optimization involving arbitrary numbers of agents. The updates of our method are obtained as the Nash equilibria of a local polymatrix approximation with a quadratic regularization, and can be computed efficiently by solving a linear system of equations. We prove local convergence of PCGD to stable fixed points for n-player general-sum games, and show that it does not require adapting the step size to the strength of the player-interactions. We use PCGD to optimize policies in multi-agent reinforcement learning and demonstrate its advantages in Snake, Markov soccer and an electricity market game. Agents trained by PCGD outperform agents trained with simultaneous gradient descent, symplectic gradient adjustment, and extragradient in Snake and Markov soccer games and on the electricity market game, PCGD trains faster than both simultaneous gradient descent and the extragradient method.