Simon Shaolei Du 杜少雷
About MeI am an assistant professor in the Paul G. Allen School of Computer Science & Engineering at University of Washington. My research interests are broadly in machine learning such as deep learning, representation learning, reinforcement learning and data selection. Prior to starting as faculty, I was a postdoc at Institute for Advanced Study of Princeton, hosted by Sanjeev Arora. I completed my Ph.D. in Machine Learning at Carnegie Mellon University, where I was co-advised by Aarti Singh and Barnabás Póczos. Previously, I studied EECS and EMS at UC Berkeley. I have also spent time at Simons Institute and research labs of Meta, Google and Microsoft. Prospective students, postdocs, and visitors, please send me an email with your CV. Selected Awards
Research Focus and Selected PublicationsRepresentation Learning Theory and Data Selection Algorithms
We studied when pretraining provably improves the performance of the downstream task. Based on our theory, we developed an active learning algorithm to select the most relevant pretraining data.
Fundamental Limits of Reinforcement Learning
We develop algorithms to obtain optimal sample complexity guarantees for reinforcement learning. In particular, we showed the sample complexity can be independent of the planning horizon.
Optimization and Generalization in Over-Parameterized Neural Networks
We proved the first set of global optimization and generalization guarantees for over-parameterized neural networks in the neural tanget kernel regime [Wikipedia]. Also see our [Blog] for a quick sumary. Recently, we found over-parameterization can exponentially slow down convergence.
Reinforcement Learning with Function Approximation
We studied the necessary and sufficient conditions that permit efficient learning for reinforcement learning problems with a large state space.
Multi-Agent Reinforcement Learning (MARL)
We initiated the study on what dataset permits solving offline reinforcement learning problems. We also study MARL with function approximation that avoids the exponential dependency on the # of agents.
|