Publications
Click any title to view the PDF. Bold indicates this author. Press / to search.
-
International Conference on Machine Learning (ICML), 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.
-
Advances in Neural Information Processing Systems (NeurIPS), Vancouver, Canada, 2024.PDF
Abstract
We conduct an asymptotic analysis of Unified Distributed SGD (UD-SGD) spanning decentralized SGD, local SGD, and Federated Learning patterns, and assess how different sampling strategies (i.i.d., shuffling, Markovian) affect convergence speed via the limiting covariance matrix. Our results extend existing theories on linear speedup and asymptotic network independence, and reveal that a few agents using highly efficient sampling can match or surpass a majority using moderately improved strategies, providing new insights beyond worst-agent analyses.
-
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 Artificial Intelligence and Statistics (AISTATS), Valencia, Spain, 2024.PDF
Abstract
We conduct an in-depth asymptotic analysis of Two-Timescale Stochastic Approximation (TTSA) under controlled Markovian noise via central limit theorem (CLT), uncovering coupled dynamics not addressed by previous CLTs that only handled martingale-difference noise. We extend efficient sampling strategies from vanilla SGD to a wider TTSA context for distributed learning, and apply the CLT to gradient-based temporal-difference (GTD) algorithms with nonlinear function approximation.
-
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.
-
Advances in Neural Information Processing Systems (NeurIPS), New Orleans, LA, 2022.PDF
Abstract
We employ the notion of efficiency ordering – well-analyzed for MCMC samplers – for SGD algorithms via Loewner ordering of asymptotic covariance matrices. We show that input sequences more efficient for MCMC sampling lead to smaller covariance for SGD in the limit, and demonstrate this for non-Markovian processes whose mixing-time bounds are intractable. Empirically the ordering extends to accelerated SGD and Adam.
-
IEEE Conference on Decision and Control (CDC), Cancun, Mexico, 2022.PDF
Abstract
We consider two SIS-type viruses competing over overlaid graphs and provide the first quantitative analysis of non-trivial coexistence equilibria, proving monotonicity in effective infection rates and giving upper bounds in terms of spectral radii $\rho(\cdot)$ of the underlying graphs and positive equilibria of single-virus systems.
-
IEEE Transactions on Mobile Computing, 2022.PDF
Abstract
We explicitly minimize file-transfer time over heterogeneous-rate Bernoulli channels in the OSA model – showing that max-throughput policy does not minimize file-transfer time in general. We give static and dynamic optimal policies, a mixed-integer programming reformulation that drastically cuts compute time, and a heuristic with online implementation under unknown channel parameters.
-
International Symposium on Modeling and Optimization in Mobile, Ad Hoc, and Wireless Networks (WiOpt), Philadelphia, PA, 2021.PDF
Abstract
We show that maximizing long-term throughput in Opportunistic Spectrum Access does not minimize file transfer time in general – especially for small files – and provide static- and dynamic-optimal policies via stochastic shortest-path formulation, with an online algorithm whose regret is bounded by $O(\sqrt{T \log T})$.
-
IEEE Vehicular Technology Conference (VTC2020-Fall), Victoria, Canada, 2020.PDF
Abstract
We analyze packet-delivery ratio and mean delay of DSRC-based V2V communication under IEEE 802.11p, and develop the Semi-persistent Contention Density Control (SpCDC) scheme. Simulations show SpCDC increases PDR by $>10\%$ and decreases mean delay by $>50\%$ in heavy-load scenarios.