Our new preprint with Yuki Takezawa, Last-Iterate Convergence Rate of Normalized Gradient Descent under Hölder Smoothness, establishes last-iterate guarantees for convex Hölder-smooth objectives, including an optimal-order rate with a linearly decreasing stepsize.
Egor Shulgin joined the team as a postdoctoral researcher in September 2026.
Our new preprint, What's in a Smoothness Constant? Tighter Rates for Local SGD with Bounded Second-Order Heterogeneity, establishes tighter convergence guarantees and nearly matching lower bounds for Local SGD on general convex objectives under bounded second-order heterogeneity.
Our paper Last Iterate Convergence of AdaGrad-Norm for Convex Non-Smooth Optimization received the Best Application Paper Award at MOTOR 2026.
My collaborators and I had several papers accepted at 2026 conferences: one at AISTATS, one at ICLR, five at ICML, two at CPAL, and two at UAI.
A new preprint studies high-probability convergence with an arbitrary clipping level.
The paper Low-Resource Machine Translation through the Lens of Personalized Federated Learning, joint with Viktor Moskvoretskii, Nazarii Tupitsa, Chris Biemann, Samuel Horvath, and Irina Nikishina, was accepted to EMNLP 2024 Findings.
Remove that Square Root: A New Efficient Scale-Invariant Version of AdaGrad.
Seven new preprints were added, including work on intermediate gradient methods, heavy-tailed stochastic minimization, Byzantine robust learning, and distributed variational inequalities.
Two papers were accepted to ICML 2023.
New preprint on convergence of proximal point and extragradient-based methods beyond monotonicity.
Presentation on variance reduction for Byzantine-robust distributed optimization.
Outstanding Reviewer Award at NeurIPS 2022.
Two papers were accepted to NeurIPS 2022.
I started as a postdoctoral researcher at MBZUAI in Abu Dhabi, in the groups of Samuel Horváth and Martin Takáč.