Jingfeng Wu

Jingfeng Wu / 吴京风

I am a Ph.D. student (2019 - now) at Johns Hopkins University, Computer Science Department. I am supervised by Prof. Vladimir Braverman. Previously I obtained B.S. (Mathematics, 2012 - 2016) and M.S. (Applied Math, 2016 - 2019) from Peking University, School of Mathematical Sciences.

Email / CV / Google Scholar / Github / Twitter

News

[Sep. 2021] Two papers are accepted to NeurIPS 2021.

[Sep. 2021] Talk on problem-dependent bound of constant-stepsize SGD at JHU CS Theory Seminar.

[May 2021] Awarded MINDS Summer Data Science Fellowship. Many thanks to the committee.

[May 2021] One paper is accepted to COLT 2021.

[Mar. 2021] Talk on the implicit bias of SGD at UCLA.

[Feb. 2021] In a relationship with Yuan, happy Valentine's day~

[Jan. 2021] One paper is accepted to ICLR 2021.

[Nov. 2020] Talk on the implicit bias of SGD at JHU CS Theory Seminar.

[Oct. 2020] One paper will be presented at Workshop on Optimization for Machine Learning (OPT).

[May 2020] Two papers get accepted to ICML 2020.

[Mar. 2020] Stay home stay safe.

[Nov. 2019] One paper will be orally presented at Workshop on Optimization for Machine Learning (OPT).

[Sep. 2019] Starting working at Hopkins. Veritas vos liberabit!

[Jun. 2019] Graduated from Peking University.

[Apr. 2019] One paper is accepted to ICML 2019.

[Mar. 2019] One paper is accepted as an oral presentation at CVPR 2019.

[Dec. 2018] Looking for a Ph.D. position in machine learning / statistical learning starting from fall, 2019.

Research

I am interested in mathematical aspects of machine learning algorithms. From the theory side, I work on understanding the algorithmic bias and the algorithmic complexity. From the application side, I develop algorithms with provable guarantees. Currently I work a lot with SGD and MDP.

	Last Iterate Risk Bounds of SGD with Decaying Stepsize for Overparameterized Linear Regression Jingfeng Wu^, Difan Zou^, Vladimir Braverman, Quanquan Gu, Sham M. Kakade arXiv, 2021 bibtex / arXiv We prove problem-dependent excess risk bounds for decaying-stepsize SGD for overparameterized linear regression problems.
	The Benefits of Implicit Regularization from SGD in Least Squares Problems Difan Zou^, Jingfeng Wu^, Vladimir Braverman, Quanquan Gu, Dean P. Foster, Sham M. Kakade Conference on Neural Information Processing Systems (NeurIPS), 2021 bibtex / arXiv We show that in a broad class of interesting least square instances, SGD is always nearly as good as ridge regression, but ridge regression could be much worse than SGD.
	Gap-Dependent Unsupervised Exploration for Reinforcement Learning Jingfeng Wu, Vladimir Braverman, Lin F. Yang ICML Workshop on Reinforcement Learning Theory, 2021 bibtex / arXiv / poster We show an improved sample complexity for unsupervised reinforcement learning when the problem instances have constant sub-optimality gap.
	Lifelong Learning with Sketched Structural Regularization Haoran Li, Aditya Krishnan, Jingfeng Wu, Soheil Kolouri, Praveen K. Pilly, Vladimir Braverman Asian Conference on Machine Learning (ACML), 2021 bibtex / arXiv We show sketching methods improve structural regularization algorithms for lifelong learning.
	Benign Overfitting of Constant-Stepsize SGD for Linear Regression Difan Zou^, Jingfeng Wu^, Vladimir Braverman, Quanquan Gu, Sham M. Kakade Annual Conference on Learning Theory (COLT), 2021 bibtex / arXiv / slides We study the generalization bounds of SGD for overparameterized linear regression problems.
	Accommodating Picky Customers: Regret Bound and Exploration Complexity for Multi-Objective Reinforcement Learning Jingfeng Wu, Vladimir Braverman, Lin F. Yang Conference on Neural Information Processing Systems (NeurIPS), 2021 bibtex / arXiv / slides / poster / code We study the regret bound and sample complexity for multi-objective reinforcement learning with potentially adversarial preferences.
	Direction Matters: On the Implicit Bias of Stochastic Gradient Descent with Moderate Learning Rate Jingfeng Wu, Difan Zou, Vladimir Braverman, Quanquan Gu International Conference on Learning Representations (ICLR), 2021 bibtex / arXiv / slides / poster We show a directional bias for SGD with moderate learning rate. This particular effect cannot be achieved by GD or SGD with small learning rate.
	Obtaining Adjustable Regularization for Free via Iterate Averaging Jingfeng Wu, Vladimir Braverman, Lin F. Yang International Conference on Machine Learning (ICML), 2020 bibtex / arXiv / slides / code We show that an l2-type regularization effect could be achieved via properly averaging an optimization path in many cases.
	On the Noisy Gradient Descent that Generalizes as SGD Jingfeng Wu, Wenqing Hu, Haoyi Xiong, Jun Huan, Vladimir Braverman, Zhanxing Zhu International Conference on Machine Learning (ICML), 2020 bibtex / arXiv / slides / code For gradient methods with noises, we show that the distribution classes of the noises do not affect their regularization effects.
	Tangent-Normal Adversarial Regularization for Semi-supervised Learning Bing Yu^, Jingfeng Wu^, Jinwen Ma, Zhanxing Zhu Conference on Computer Vision and Pattern Recognition (CVPR), 2019, oral bibtex / arXiv / slides / poster / code We present a novel manifold regularization method for semi-supervised learning. The regularizer is realized via adversarial training.
	The Anisotropic Noise in Stochastic Gradient Descent: Its Behavior of Escaping from Minima and Regularization Effects Zhanxing Zhu^, Jingfeng Wu^, Bing Yu, Lei Wu, Jinwen Ma International Conference on Machine Learning (ICML), 2019 bibtex / arXiv / slides / poster / code We study the noise structure of stochastic gradient descent, and demonstrate its benefits on helping the dynamic escape from sharp minima.