Faculty
Dr. Sebastian Stich ist seit dem 1. Dezember 2021 Tenure-Track-Faculty am CISPA Helmholtz-Zentrum für Informationssicherheit. Seit Juni 2020 ist er Mitglied des European Lab for Learning and Intelligent Systems. Vom 1. Dezember 2016 bis zum 30. November 2021 war er Forscher an der EPFL unter der Leitung von Prof. Martin Jaggi im Machine Learning and Optimization Laboratory (MLO). Vom 1. November 2014 bis 31. Oktober 2016 arbeitete er mit Prof. Yurii Nesterov und Prof. François Glineur am Center for Operations Research and Econometrics (CORE) und am ICTEAM. Vom 15. September 2010 bis 30. September 2014 war er Doktorand in der Forschungsgruppe von Prof. Emo Welzl, betreut von Prof. Bernd Gärtner und Christian Lorenz Müller. Von September 2005 bis März 2010 absolvierte er seinen Bachelor und Master in Mathematik an der ETH Zürich.
International Conference on Machine Learning (ICML)
Federated Optimization with Doubly Regularized Drift Correction
International Conference on Machine Learning (ICML)
Spectral Preconditioning for Gradient Methods on Graded Non-convex Functions
International Conference on Machine Learning (ICML)
On Convergence of Incremental Gradient for Non-convex Smooth Functions
International Conference on Machine Learning (ICML)
Non-Convex Stochastic Composite Optimization with Polyak Momentum
Conference on Learning Theory (COLT)
The Limits and Potentials of Local SGD for Distributed Heterogeneous Learning with Intermittent Communication
International Conference on Learning Representations (ICLR)
An improved analysis of per-sample and per-update clipping in federated learning
International Conference on Learning Representations (ICLR)
Communication-Efficient Gradient Descent-Accent Methods for Distributed Variational Inequalities: Unified Analysis and Local Updates
International Conference on Learning Representations (ICLR)
EControl: Fast Distributed Optimization with Compression and Error Control.
Optimization Methods and SoftwareDecentralized gradient tracking with local steps
Conference on Neural Information Processing Systems (NeurIPS)
Adaptive SGD with Polyak stepsize and Line-search: Robust Convergence and Variance Reduction