The laboratory staff prepared two articles that were submitted to the world's leading scientific conference on machine learning ICML 2023. The event was held from July 23 to 29, 2023 in Hawaii with the possibility of online participation.
The paper Is Consensus Acceleration Possible in Decentralized Optimization over Slowly Time-Varying Networks? prepared by Dmitry Metelev, Alexander Rogozin, Dmitry Kovalev and Alexander Gasnikov.
«Our work was on the list of the past, it is of purely theoretical interest. The article discusses the lower bounds in a distributed optimization problem on a slowly changing network, as well as specific limitations on it, under which it is possible to obtain theoretical acceleration of distributed algorithms,» Dmitry Metelev said.
The paper High-Probability Bounds for Stochastic Optimization and Variational Inequalities: the Case of Unbounded Variation was prepared by Abdurakhmon Sadiev, Marina Danilova, Eduard Gorbunov, Samuel Horvath, Gauthier Gidel, Pavel Dvurechensky, Alexander Gasnikov and Peter Richtarik. The key objective of the study was to analyze the convergence of clipping methods for the minimization problem and variational inequalities under more general noise assumptions.
«I saw the report by Eduard [Gorubnov] on the Optimization and Statistics Youtube channel. I wrote to Eduard, and then we agreed to work together on the paper. The coordinator of the work on the project was Eduard himself. I started studying previous papers and eventually proved the main lemma, then conducted a convergence analysis of clipped-SGDA,» Abdurakhmon Sadiev shared his impressions of the work.
The conference presented cutting-edge research on all aspects of machine learning used in closely related fields, in particular artificial intelligence, statistics and data science, as well as in important applied fields such as machine vision, computational biology, speech recognition and robotics. The participants were academic and industrial researchers, entrepreneurs, engineers, graduate students and other leading specialists.