The First Optimal Distributed SGD in the Presence of Data, Compute and Communication Heterogeneity

Explore cutting-edge research on parallel and distributed optimization methods in this 29-minute conference talk from the "One World Optimization Seminar in Vienna" workshop. Delve into the complexities of designing efficient algorithms for parallel optimization, particularly in heterogeneous computing environments. Discover the groundbreaking work on establishing optimal time complexities for parallel optimization methods, including Rennala SGD and Malenia SGD, which address compute heterogeneity with unbiased stochastic gradient oracles. Learn about the novel Shadowheart SGD algorithm, which tackles both compute and communication heterogeneity. Examine the surprising implications for asynchronous optimization methods and how these findings challenge previous approaches. Gain insights into the lower bounds and optimal algorithms developed for both data homogeneous and heterogeneous regimes. Understand the significance of alternating fast asynchronous computation with infrequent synchronous update steps in achieving optimal performance.

Peter Richtarik - The First Optimal Distributed SGD in the Presence of Data, Compute...