I am an Assistant Professor in the Department of Computer Science and Engineering at Oakland University, based in Rochester, MI. My research focuses on machine learning theory, distributed optimization, and graph sampling. In particular, I design efficient algorithms grounded in applied probability and Markov chain theory — pushing the boundaries of how quickly and effectively learning tasks can be performed over networks.
News
- Sep 2026 Our paper Geometry-Aware Score-Repellent Monte Carlo was accepted at NeurIPS 2026! ↗
- Apr 2026 Our paper Score-Repellent Monte Carlo: Toward Efficient Non-Markovian Sampler with Constant Memory in General State Spaces was accepted at ICML 2026 as a Spotlight (top 2.2% of 23,918 submissions)! ↗
- Feb 2026 Awarded the 2026 URC Faculty Research Fellowship from Oakland University to support research on efficient sampling and distributed learning.
Selected Publications
-
Advances in Neural Information Processing Systems (NeurIPS), Sydney, Australia, 2026.Poster PDF
Abstract
The recently proposed Score-Repellent Monte Carlo (SRMC) method improves MCMC sampling by tilting the target density $\pi$ in $\mathbb{R}^d$ with the factor $\exp(-\alpha\,\theta_n^\top s(x))$, where the score function $s(x) = \nabla_x \log \pi(x)$, and $\theta_n$ is the score average from past samples. However, a single scalar repellence strength $\alpha\ge0$ cannot handle both the steep and the flat directions in anisotropic energy landscapes at once. We introduce Geometry-Aware SRMC (GA-SRMC), which generalizes the Euclidean alignment to $\theta^\top M s(x)$ for any symmetric positive-definite matrix $M$. A stochastic-approximation central limit theorem under a trace constraint on $M$ identifies the trace-normalized inverse Fisher $M_{\rm opt}\propto S^{-1}$ as the unique minimizer of the worst-direction covariance bound, elevating the Fisher information $S=\text{Cov}_\pi(s,s)$ from a proposal-side preconditioner (e.g., FisherMALA, natural gradient) to the bound-optimal repulsion-side matrix, with a single online estimate $\widehat S_n$ serving both roles at no additional computational cost. The full-Fisher online metric $M_n$ converges almost surely to $M_{\rm opt}$ and satisfies a CLT, providing a theoretical guarantee for the online implementation of the GA-SRMC framework. On an $8$-target simulation spanning the condition number $\kappa\in[1.3,100]$ and dimension $d\in[2,785]$, GA-SRMC dominates SRMC in ESS by $\boldsymbol{4.3\times}$ on anisotropic Gaussian ($\kappa=100$) and $\boldsymbol{6.8\times}$ on MNIST under the MALA baseline, by $\boldsymbol{16\text{–}32\%}$ on three well-conditioned logistic-regression posteriors, and by up to $\boldsymbol{+48\%}$ under FisherMALA; gains peak on the most anisotropic targets, and posterior-mean MSE is lower than with SRMC in all $16$ full-Fisher implementations.
-
International Conference on Machine Learning (ICML), Seoul, Korea, 2026.Spotlight PDF
Abstract
We propose a score-repellent Monte Carlo sampler that is non-Markovian yet requires only $O(1)$ memory, generalizing self-repellent ideas from discrete graphs to general state spaces. The sampler shapes its proposal distribution using a running score estimate, achieving asymptotically minimal sampling variance without storing the full history.
-
International Conference on Machine Learning (ICML), Vancouver, Canada, 2025.Oral PDF
Abstract
We propose a history-driven target (HDT) framework in MCMC to improve any random walk algorithm on discrete state spaces, such as general undirected graphs, for efficient sampling from a target distribution $\boldsymbol{\mu}$. Recent innovations like the self-repellent random walk (SRRW) achieve near-zero variance by prioritizing under-sampled states through transition-kernel modifications, but suffer high computational overhead and require time-reversibility. HDT instead introduces a history-dependent target $\boldsymbol{\pi}[\mathbf{x}]$ to replace the original target $\boldsymbol{\mu}$, where $\mathbf{x}$ is the empirical measure of past visits. This preserves a lightweight implementation, achieves compatibility with both reversible and non-reversible MCMC samplers, and retains unbiased samples with near-zero variance.
-
International Conference on Learning Representations (ICLR), Vienna, Austria, 2024.Oral PDF
Abstract
We replace the standard linear Markovian token in random-walk-based distributed SGD by a Self-Repellent Random Walk (SRRW) – a nonlinear Markov chain parameterized by $\alpha > 0$ that is less likely to revisit recently visited states. Our SA-SRRW algorithm achieves almost-sure convergence and a CLT, with asymptotic covariance always smaller than that of an algorithm driven by the base Markov chain and decreasing at rate $O(1/\alpha^2)$ – amplifying SRRW’s benefit in the stochastic optimization context.
-
International Conference on Machine Learning (ICML), Honolulu, HI, 2023.🏆 Outstanding Paper Award PDF
Abstract
We design a self-repellent random walk (SRRW) parameterized by $\alpha > 0$ that is less likely to transition to highly visited nodes. We prove almost-sure convergence of the empirical distribution to the target stationary distribution and a central limit theorem showing that stronger repellence (larger $\alpha$) yields strictly smaller asymptotic covariance in the Loewner order. For SRRW-driven MCMC, the decrease in asymptotic sampling variance is $O(1/\alpha)$, eventually going to zero.
Students
- Jingpei Zhong PhD student Since Fall 2026
- Shubham Jha PhD student Since Fall 2026
- Tasmin Karim PhD student Since Fall 2026
Teaching
- CSI 2560 Computational Linear Algebra Winter 2026, Winter 2027
- CSI 2470 Introduction to Computer Networks Fall 2025, Fall 2026