Yuanzhi Li

I'm a Assistant Professor in the Machine Learning Department at CMU. Previously I was a postdoc at Stanford Computer Science Department (20180-2019). I obtained my Ph.D in computer science at Princeton University (2014-2018) and my B.S.E. in computer science and mathematics at Tsinghua University (2010-2014).

I work on Deep Learning (Theory).

Contact

$l at andrew dot cmu dot edu, replace $ with my first name.

Papers (See google scholar for the full list)

  • A Convergence Theory for Deep Learning via Over-Parameterization. With Zeyuan Allen-Zhu and Zhao Song. Manuscript
  • Learning and Generalization in Overparameterized Neural Networks, Going Beyond Two Layers. With Zeyuan Allen-Zhu and Yingyu Liang. Submitted
  • On the Convergence Rate of Training Recurrent Neural Networks. With Zeyuan Allen-Zhu and Zhao Song. Submitted
  • Chasing Nested Convex Bodies Nearly Optimally. With Sébastien Bubeck, Yin Tat Lee and Mark Sellke. Submitted
  • Competitively Chasing Convex Bodies. With Sébastien Bubeck, Yin Tat Lee and Mark Sellke. Submitted
  • Algorithmic Framework for Model-based Deep Reinforcement Learning with Theoretical Guarantees. With Yuping Luo, Huazhe Xu, Yuandong Tian, Trevor Darrell and Tengyu Ma. Submitted
  • Learning Overparameterized Neural Networks via Stochastic Gradient Descent on Structured Data. With Yingyu Liang. NIPS 2018
  • Online Improper Learning with an Approximation Oracle. With Elad Hazan, Wei Hu and Zhiyuan Li. NIPS 2018
  • Neon2: Finding Local Minima via First-Order Oracles. With Zeyuan Allen-Zhu. NIPS 2018
  • Algorithmic Regularization in Over-parameterized Matrix Recovery. With Tengyu Ma and Hongyang Zhang. COLT 2018
  • Learning Mixtures of Linear Regressions with Nearly Optimal Complexity. With Yingyu Liang. COLT 2018
  • The Well Tempered Lasso. With Yoram Singer. ICML 2018
  • Make the Minority Great Again: First-Order Regret Bound for Contextual Bandits. With Zeyuan Allen-Zhu and Sebastien Bubeck. ICML 2018
  • An Alternative View: When Does SGD Escape Local Minima?. With Robert Kleinberg and Yang Yuan. ICML 2018
  • Operator Scaling via Geodesically Convex Optimization, Invariant Theory and Polynomial Identity Testing. With Zeyuan Allen-Zhu, Ankit Garg, Rafael Oliveira and Avi Wigderson. STOC 2018
  • An homotopy method for Lp regression provably beyond self-concordance and in input-sparsity time. With Sebastien Bubeck, Michael B. Cohen and Yin Tat Lee. STOC 2018
  • Linear algebraic structure of word senses, with applications to polysemy. With Sanjeev Arora, Yingyu Liang, Tengyu Ma and Andrej Risteski. TACL 2018
  • Sparsity, variance and curvature in multi-armed bandits. With Sebastien Bubeck and Michael B. Cohen. ALT 2018
  • An Instance Optimal Algorithm for Top-k Ranking under the Multinomial Logit Model. With Xi Chen and Jieming Mao. SODA 2018
  • Linear Convergence of a Frank-Wolfe Type Algorithm over Trace-Norm Balls. With Zeyuan Allen-Zhu, Elad Hazan and Wei Hu. NIPS 2017
  • Convergence Analysis of Two-layer Neural Networks with ReLU Activation . With Yang Yuan. NIPS 2017
  • Much Faster Algorithms for Matrix Scaling. With Zeyuan Allen-Zhu, Rafael Oliveira and Avi Wigderson. FOCS 2017
  • Fast Global Convergence of Online PCA. With Zeyuan Allen-Zhu. FOCS 2017
  • Near-Optimal Design of Experiments via Regret Minimization. With Zeyuan Allen-Zhu, Aarti Singh and Yining Wang. ICML 2017
  • Provable Alternating Gradient Descent for Non-negative Matrix Factorization with Strong Correlations. With Yingyu Liang. ICML 2017
  • Follow the Compressed Leader: Faster Algorithm for Matrix Multiplicative Weight Updates. With Zeyuan Allen-Zhu. ICML 2017
  • Faster Principal Component Regression via Optimal Polynomial Approximation to sgn(x). With Zeyuan Allen-Zhu. ICML 2017
  • Doubly Accelerated Methods for Faster CCA and Generalized Eigendecomposition. With Zeyuan Allen-Zhu. ICML 2017
  • RAND-WALK: a latent variable model approach to word embeddings. With Sanjeev Arora, Yingyu Liang, Tengyu Ma and Andrej Risteski. TACL 2016
  • Even Faster SVD Decomposition Yet Without Agonizing Pain. With Zeyuan Allen-Zhu. NIPS 2016
  • Approximate maximum entropy principles via Goemans-Williamson with applications to provable variational methods. With Andrej Risteski. NIPS 2016
  • Tight algorithms and lower bounds for approximately convex optimization. With Andrej Risteski. NIPS 2016
  • Recovery Guarantee of Non-negative Matrix Factorization via Alternating Updates. With Yingyu Liang and Andrej Risteski. NIPS 2016
  • Recovery guarantee of weighted low-rank approximation via alternating minimization. With Yingyu Liang and Andrej Risteski. ICML 2016
  • A Theoretical Analysis of NDCG Ranking Measures. With Yining Wang, Liwei Wang, Di He, Wei Chen and Tie-Yan Liu. COLT 2013
  • ...

    • Teaching

  • COS423, Princeton: Theory of Algorithms, Fall 2016. TA
  • COS521, Princeton: Advance algorithm design, Spring 2016. TA
  • COS10725, CMU: Optimization in Machine Learning (Convex Optimization), Spring and Fall, 2020-2022. Instructor