Jason Lee

Timeline

1. By 2015

   Mathematics at Duke and Stanford

   A native of Cupertino, California, he earns a BS in mathematics at Duke University with Mauro Maggioni as his advisor. In 2015, he completes a PhD in computational and applied mathematics at Stanford University, advised by Trevor Hastie and Jonathan Taylor.

2. 2016

   Gradient descent reaches local minima

   In 2016, he coauthors Gradient Descent Only Converges to Minimizers with Max Simchowitz, Michael I. Jordan, and Benjamin Recht. The paper uses the Stable Manifold Theorem to show that gradient descent converges to a local minimum with probability one from a random starting point.

3. By 2018

   Assistant professor at USC

   By 2018, he is an assistant professor in Data Sciences and Operations at the University of Southern California. Before USC, he works as a postdoctoral researcher at UC Berkeley with Michael Jordan.

4. 2019

   Training deep networks to global minima

   With Simon Du, Haochuan Li, Liwei Wang, and Xiyu Zhai, he coauthors Gradient Descent Finds Global Minima of Deep Neural Networks in 2019. They prove that gradient descent reaches zero training loss in polynomial time for a deep neural network with extra parameters and residual connections.

5. 2021

   Guarantees for policy gradient methods

   In 2021, he coauthors On the Theory of Policy Gradient Methods: Optimality, Approximation, and Distribution Shift with Alekh Agarwal, Sham M. Kakade, and Gaurav Mahajan. The work gives mathematical guarantees for the computation, approximation, and sample requirements of these reinforcement learning methods in discounted Markov decision processes.

6. 2023

   Samsung AI Researcher of the Year

   In 2023, he receives a Samsung AI Researcher of the Year award as a Princeton University professor. He is one of five researchers to receive the award.

7. 2024

   Medusa speeds language model generation

   In 2024, he coauthors Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads. Medusa adds decoding heads that predict several subsequent tokens in parallel, then checks candidate continuations together to reduce the number of generation steps.

8. 2025

   Associate professor at Berkeley

   In 2025, he joins UC Berkeley as an associate professor of statistics with a joint appointment in Electrical Engineering and Computer Sciences. His earlier roles include associate professor at Princeton, research scientist at Google DeepMind, and member of the Institute for Advanced Study.

9. As of October 2026

   Senior scientist at Simons

   As of October 2026, he is an associate professor in UC Berkeley's Departments of Electrical Engineering and Computer Sciences and Statistics, and Senior Scientist at the Simons Institute for the Theory of Computing. His research includes the foundations of deep learning, representation learning, and reinforcement learning.