My work primarily focuses on the statistical and computational aspects of machine learning, signal processing, and data science. Currently, I am particularly interested in high-dimensional statistics and random matrix theory, as well as their interactions with (deep or not so deep) neural networks.
For more information, see here for my headshot and here for a short bio in Chinese.
I am always looking for self-motivated interns, masters, and Ph.D. students, with a strong background and genuine interest in statistics and machine learning theory and methods on (but not limited to) the following topics:
theoretical foundations and methods of deep learning, model compression, and efficient computation (e.g., randomized numerical linear algebra, stochastic and/or distributed optimization).
If you are passionate about these areas and eager to work in a collaborative and research-driven environment, feel free to reach out.
[May 2025] Our work on the statistical inversion bias of random sampling, with application to sub-sampled Newton methods, has been selected as an Oral paper at ICML 2025!
In this paper, we show that commonly used random sampling techniques ranging from uniform and leverage, to Hadamard transform-based, all exhibit statistical inversion bias.
And, this bias can be effectively corrected (sometimes without additional computational burden) using RMT.
This work is led by my PhD student Chengmei Niu—check out our preprint here!
[Apr 2025] I will be serving as an Area Chair for NeurIPS 2025.
Check out the work-in-progress lecture notes on “Random Matrix Theory for Modern Machine Learning.”
Feedback and suggestions are very welcome—feel free to share your thoughts!
[Jan 2005] Our work on the breakdown of Gaussian Universality for large-dimensional (convex) generalized linear classifier accepted at ICLR'2025, check out our preprint here!
[Dec 2024] I will be serving as an Area Chair for IJCNN 2025.
[Nov 2024] I will be serving as an Area Chair for ICML 2025.
[Aug 2024] I will be serving as an Area Chair for ICLR 2025.
[Jun 2024] One paper at EUSIPCO'2024 on a novel and efficient RMT-improved Direction of Arrival estimation method. We show that standard ESPRIT is biased in the case of large arrays, but can be effectively de-biased using RMT. The paper is listed as a Best Student Paper Candidate of EUSIPCO'2024! Check the extended version here.
[May 2024] One paper at ICML'2024 on RMT for Deep Equilibrium Model (DEQ, a typical implicit NN model) that provides very explicit connections between implicit and explicit NNs. Given a DEQ, check our preprint here for the recipe to design an “equivalent” simple explicit NN!
[Aug 2022] One paper at NeurIPS'2022 on the eigenspectral structure of Neural Tangent Kernel (NTK) of fully-connected Deep neural networks for Gaussian mixture input data, with a compelling application to “lossless” sparsification and quantization of DNN models! This extends our previous paper at ICLR'2022. See more details here.
[Jun 2022] Our book “Random Matrix Methods for Machine Learning” with Cambridge University Press, see more details here!
