Junwei Lu

We know optimism is provably efficient for online RL. What about offline RL? It turns out simply flipping the sign of the bonus is minimax optimal! Given a dataset, pessimism is the best effort we can make. arxiv.org/abs/2012.15085 Just leave pessimism to 2020. Happy new year~!

Struggling with fine-tuning BERT models? Overfit your tasks again? Check our recent work on ACL 2020 "SMART: Robust and Efficient Fine-Tuning for Pre-trained Natural Language Models through Principled Regularized Optimization" arxiv.org/abs/1911.03437

Asst Professor @lu_junwei is organizing an online seminar on Mathematical Foundations of Data Science w. @EthanFangXY, Niao He, @zhaoran_wang, Zhuoran Yang, and @tourzhao. Next speaker: @UMengineering Prof Alfred Hero May 12 @ 3pm EST via Zoom. More info: hsph.me/28s