News
Papers, talks, and announcements, newest first.
New playground and note on decentralized backtracking: “An accelerated primal–dual algorithm with backtracking for decentralized constrained optimization” (with Q. Xu and N. S. Aybat), in the Journal of Nonlinear and Variational Analysis special issue in honor of Yurii Nesterov. Agents on a network find their own stepsizes by a local test, synchronized by one max-consensus per iteration. Details
New preprint: “DREAM: delay-resistant acceleration with momentum” (with M. H. Rafi, N. S. Aybat, and W. U. Bajwa). Neither heavy ball nor Nesterov’s method can be retuned to stay accelerated when gradients arrive late; a momentum method with delay-matched extrapolation can, and it meets the delayed lower bound up to logarithms. Details
Talk at the 2026 Annual Meeting of the New York–New Jersey–Pennsylvania Section of SIAM (SIAM NNP26), Rutgers University, Piscataway (October 25). Details
“Privacy of SGD under Gaussian or heavy-tailed noise: guarantees without gradient clipping” (with U. Şimşekli, S. Yıldırım, and L. Zhu) is accepted for publication in the Journal of Machine Learning Research. Details
We will be holding a ROADS workshop, “Advances in Large-Scale and Stochastic Optimization,” on May 21–22, 2027, at Rutgers Business School, organized with Jonathan Eckstein, Xiaodong Lin, and Andrzej Ruszczyński. Details
Invited talk, “Robust and Risk-Sensitive Acceleration in Gradient Methods,” at the Financial/Actuarial Mathematics Seminar, University of Michigan, Ann Arbor (October 7). Details
On sabbatical leave for 2026–27 as a Visiting Professor at Columbia University, with the Department of Industrial Engineering and Operations Research and the Data Science Institute.
“RESIST: resilient decentralized learning using consensus gradient descent” (with C. Fang, R. Dixit, and W. U. Bajwa) appears in Transactions on Machine Learning Research with a Featured certification. Details
“Accelerated gradient methods with biased gradient estimates: risk sensitivity, high-probability guarantees, and large deviation bounds” (with Y. Syed and N. S. Aybat) appears in the Journal of Nonlinear and Variational Analysis. Details
“Mean-semideviation-based distributionally robust learning with weakly convex losses” (with L. Zhu and A. Ruszczyński) appears in Mathematical Programming 215(1). Details
“Robustly stable accelerated momentum methods with a near-optimal L2 gain and H∞ performance” appears in Mathematics of Operations Research. The paper is dedicated to Michael L. Overton. Details
“Entropic risk-averse generalized momentum methods” (with B. Can) appears in Optimization Methods and Software 40(6). Details
Two papers appear in the Journal of Machine Learning Research — on constrained sampling with penalized Langevin Monte Carlo, and on high-probability, risk-averse guarantees for stochastic accelerated primal-dual methods — alongside “Robust accelerated primal-dual methods for computing saddle points” in the SIAM Journal on Optimization and a NeurIPS 2024 paper on stochastic non-convex minimax optimization. Details