Jiaming Yang
|
Jiaming Yang (杨佳明)
Email: jiamyang [at] umich [dot] edu |
About Me
I am on the job market and expect to graduate in May 2027.
I am a fifth-year Ph.D. student in Computer Science at the University of Michigan, where I am advised by Prof. Michał Dereziński. Prior to that, I received my bachelor's degree from the School of Mathematical Sciences at Peking University.
My current research focuses on randomized numerical linear algebra (RandNLA), with applications to large-scale optimization and machine learning. More broadly, I am interested in machine learning, statistics, and optimization, with a focus on their theoretical foundations.
Outside of research, I enjoy rock climbing (both indoor and outdoor), photography, and playing poker.
Research Interests
Randomized Numerical Linear Algebra (RandNLA)
Machine Learning Theory
Large-Scale Optimization
Matrix-Based Optimizers for LLM Training (e.g., Muon, Shampoo)
Publications
α-β indicates alphabetical author order.
-
Acc-RSRN: Accelerated Randomized Subspace Regularized Newton
J. Yang, P.-L. Poirion, A. Takeda, M. Dereziński
Under review, 2026
We accelerate randomized subspace Newton methods using sharper spectral-tail analysis and Nesterov momentum, achieving faster local convergence for smooth, possibly nonconvex optimization. -
Have ASkotch: A Neat Solution for Large-scale Kernel Ridge Regression
P. Rathore, Z. Frangella, J. Yang, M. Dereziński, M. Udell
Journal of Machine Learning Research (JMLR), 2026
[JMLR][arXiv]A scalable, accelerated iterative solver for full kernel ridge regression, avoiding the accuracy loss of inducing-point approximations. It provably converges linearly and outperforms state-of-the-art full and inducing-point KRR solvers on 23 large-scale tasks. -
Randomized Kaczmarz Methods with Beyond-Krylov Convergence
(α-β) M. Dereziński, D. Needell, E. Rebrova, J. Yang
SIAM Journal on Matrix Analysis and Applications, 2025
[SIMAX][arXiv]An accelerated randomized block Kaczmarz method that exploits outlying singular values to achieve Krylov-style convergence, capturing large outliers provably faster than popular Krylov methods. Its PSD variant, CD++, is competitive with CG and GMRES on benchmarks. -
Faster Linear Systems and Matrix Norm Approximation via Multi-Level Sketched Preconditioning
(α-β) M. Dereziński, C. Musco, J. Yang
Proceedings of the 2025 Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), 2025
[SODA][arXiv]Builds a Nyström preconditioner via sparse sketching, then inverts it quickly with further levels of sketching and preconditioning. This gives faster algorithms for linear systems with a few large outlying singular values, regularized systems from Gaussian process regression, and nuclear/Schatten-norm approximation. -
Solving Dense Linear Systems Faster than via Preconditioning
(α-β) M. Dereziński, J. Yang
Proceedings of the 56th Annual ACM Symposium on Theory of Computing (STOC), 2024
[STOC][arXiv]A stochastic solver for dense \(n \times n\) linear systems running in \(\tilde{O}(n^2 + nk^{\omega-1})\) time, where \(k\) is the number of large singular values (\(\tilde{O}(n^2)\) when \(k = O(n^{0.729})\)). For flat-tailed spectra, it beats both direct solvers and preconditioned CG. -
HERTA: A High-Efficiency and Rigorous Training Algorithm for Unfolded Graph Neural Networks
Y. Yang, J. Yang, W. Hu, M. Dereziński
Mathematics of Modern Machine Learning (M3L, NeurIPS), 2024
[arXiv]Trains Unfolded GNNs with a nearly-linear worst-case time guarantee while converging to the optimum of the original model, preserving its interpretability. As a byproduct, it introduces a new spectral sparsifier for normalized and regularized graph Laplacians. -
Federated Adversarial Learning: A Framework with Convergence Analysis
(α-β) X. Li, Z. Song, J. Yang
International Conference on Machine Learning (ICML), 2023
[ICML][arXiv]Proposes federated adversarial learning, cast as a min-max problem. Proves convergence in the over-parameterized regime despite multiple local updates and non-IID clients. -
Pixelated Butterfly: Simple and Efficient Sparse Training for Neural Network Models
B. Chen, T. Dao, K. Liang, J. Yang, Z. Song, A. Rudra, C. Ré
International Conference on Learning Representations (ICLR), 2022, Spotlight
[ICLR][arXiv]Sparsifies attention and MLP layers with a fixed, hardware-friendly pattern of flat block butterfly plus low-rank matrices. Trains MLP-Mixer, ViT and GPT-2 medium up to 2.5× faster than dense models on ImageNet and WikiText-103 with no drop in accuracy.
Talks
Faster Randomized Subspace Newton via Nesterov Acceleration. INFORMS Annual Meeting, San Francisco, Nov 2026
Randomized Kaczmarz Methods with Beyond-Krylov Convergence. ICCOPT, Los Angeles, Jul 2025
Faster Linear Systems via Multi-level Sketched Preconditioning. SODA, New Orleans, Jan 2025
Solving Dense Linear Systems Faster than via Preconditioning. UMich Theory Lunch, Mar 2024
Teaching
-
EECS 376: Foundations of Computer Science, University of Michigan
Graduate Student Instructor (GSI), Winter 2026
Service
Conference Reviewer: NeurIPS (2024, 2025), ICML (2025, 2026), ICLR (2027), AISTATS (2025, 2026)
Workshop Reviewer: M3L (2024), MOSS (2025)