About
I work on the foundations and algorithms of machine learning, with optimization as a central perspective. My research asks how learning systems should be designed when they are not trained once and in one place, but learn collectively from distributed, heterogeneous data and keep adapting after deployment.
My group studies optimization for modern machine learning, scalable and distributed learning, collaborative learning across models and institutions, and continual learning and machine unlearning. We combine mathematical analysis, algorithm design and empirical investigation.
Before joining CISPA I was a research scientist at EPFL and an SNSF postdoctoral fellow at UCLouvain, and I hold a PhD from ETH Zurich. I am a member of ELLIS and of the Saarbrücken ELLIS unit.
- ERC Consolidator GrantCollectiveMinds, 2025–2030
- Charles Broyden PrizeBest paper in Optimization Methods and Software (2023), awarded 2026
- Simons Institute, UC BerkeleyLong-term participant, Federated and Collaborative Learning, 2026
- Google Research Scholar Award2023
- Meta Research AwardPrivacy-Enhancing Technologies, 2022
Research
My research develops principles and algorithms for learning under computational, communication and information constraints. A recurring question is what a learning system needs to exchange, and what it needs to recompute, when data and knowledge are spread across machines, institutions or models, and when the information available to it changes over time.
- Optimization for machine learning: understanding and designing stochastic, adaptive and structured optimization methods for modern machine learning.
- Scalable and distributed learning: learning efficiently when data and computation are distributed, with an emphasis on communication, local computation, decentralization and heterogeneity.
- Collaborative and personalized learning: enabling models to share knowledge while retaining specialization across different data sources, tasks and model architectures.
- Continual learning and machine unlearning: developing principled ways to update models as information arrives, changes or must be removed.
My ERC Consolidator project CollectiveMinds develops the third direction further, studying how independently trained and heterogeneous models can exchange and combine knowledge without requiring centralized data or identical architectures.
Privacy, robustness and statistical heterogeneity arise as cross-cutting constraints across several of these directions.
Selected publications
-
NeurIPS2026Forgetting Has Neighbors: Localized Collateral Forgetting in Machine Unlearning
In Advances in Neural Information Processing Systems (NeurIPS), 2026
-
ICML2026Enhancing LLM Training via Spectral Clipping
In International Conference on Machine Learning (ICML), 2026
-
COLT2026On the Stability of Nonlinear Dynamics in GD and SGD: Beyond Quadratic Potentials
In Conference on Learning Theory (COLT), 2026
-
ICLR2026Non-Convex Federated Optimization under Cost-Aware Client Selection
In International Conference on Learning Representations (ICLR), 2026 Oral, top 1%
-
ICLR2026Composite Optimization with Error Feedback: the Dual Averaging Approach
In International Conference on Learning Representations (ICLR), 2026
-
COLT2024The Limits and Potentials of Local SGD for Distributed Heterogeneous Learning with Intermittent Communication
In Conference on Learning Theory (COLT), 2024
-
OMS2023Stochastic distributed learning with gradient quantization and double-variance reduction
Optimization Methods and Software 38(1), 91–106, 2023 Charles Broyden Prize
-
FnT2021Advances and Open Problems in Federated Learning
Foundations and Trends in Machine Learning 14(1-2), 1–210, 2021
-
JMLR2020The Error-Feedback Framework: Better Rates for SGD with Delayed Gradients and Compressed Communication
Journal of Machine Learning Research 21(237), 1–36, 2020
-
ICML2020SCAFFOLD: Stochastic Controlled Averaging for Federated Learning
In International Conference on Machine Learning (ICML), 2020
-
ICML2020A Unified Theory of Decentralized SGD with Changing Topology and Local Updates
In International Conference on Machine Learning (ICML), 2020
-
NeurIPS2020Ensemble Distillation for Robust Model Fusion in Federated Learning
In Advances in Neural Information Processing Systems (NeurIPS), 2020
-
ICLR2019Local SGD Converges Fast and Communicates Little
In International Conference on Learning Representations (ICLR), 2019
-
ICML2019Error Feedback Fixes SignSGD and other Gradient Compression Schemes
In International Conference on Machine Learning (ICML), 2019
-
ICML2019Decentralized Stochastic Optimization and Gossip Algorithms with Compressed Communication
In International Conference on Machine Learning (ICML), 2019
-
NeurIPS2018Sparsified SGD with Memory
In Advances in Neural Information Processing Systems (NeurIPS), 2018
The complete list is on Google Scholar. Work with my group at CISPA is collected on mlolab.org.
Group
I lead the Machine Learning & Optimization group at CISPA, bringing together researchers working on optimization, distributed and collaborative learning, and the foundations of modern machine learning.
Postdocs
- Anton Rodomanov since 2023
- Rotem Mulayoff since 2024
- Slavomír Hanzely since 2026
PhD students
- Xiaowen Jiang since 2023
- Yuan Gao since 2023
- Polina Dolgova since 2025
- Ali Zindari since 2025
- Akansh Maurya since 2026
- Parham Yazdkhasti since 2026
- Sharannya Ghosh since 2026
Former PhD students
- Anastasia Koloskova PhD 2023, EPFL, co-advised with Martin Jaggi. Now Assistant Professor, University of Zurich
Teaching
At Saarland University I teach advanced courses on optimization for machine learning, modern optimization methods, and game-theoretic aspects of machine learning, together with research seminars in optimization and machine learning.
- SS22–24, SS26Optimization for Machine LearningAdvanced lecture: optimization for large datasets and distributed learning, with an implementation project.
- SS25, SS27Lectures on Modern Optimization MethodsAdvanced lecture, co-taught with Anton Rodomanov; offered as an intensive two-week course.
- WS23Games in Machine LearningAdvanced lecture, co-taught with Tatjana Chavdarova.
- WS21, 22, 24, 25Topics in Optimization for Machine LearningSeminar.
- SS23Optimization Methods for Large-Scale Machine LearningPro-seminar.
Selected talks
- Oct 2026Beyond Iteration Complexity: Communication-Efficient Distributed OptimizationInvited, Colloquium in honour of Yurii Nesterov, UCLouvain, Belgium.
- Aug 2026Beyond Iteration Complexity: Communication-Efficient Distributed OptimizationInvited, Swiss Optimization Symposium, Ascona, Switzerland.
- Jan 2026Exploiting Similarity in Federated LearningInvited, Simons Institute programme on Federated and Collaborative Learning, UC Berkeley, USA.
- Nov 2025Exploiting Similarity in Federated OptimizationInvited, Workshop on Distributed Training in the Era of Large Models, KAUST, Saudi Arabia.
- Mar 2025A Universal Framework for Federated (Convex) OptimizationInvited, OIST Machine Learning Workshop, Okinawa, Japan.
- Nov 2024Privacy in Federated LearningCISPA and INRIA kick-off workshop, Campus Cyber, Paris, France.
- Aug 2024Keynote: Local Update Methods in Federated OptimizationFedKDD workshop at SIGKDD, Barcelona, Spain.
- Jun 2024Plenary: A Universal Framework for (Convex) Federated LearningEUROPT Conference on Advances in Continuous Optimization, Lund, Sweden.
- Feb 2024On Gradient Methods for Non-Convex OptimizationInvited, Symposium on Sparsity and Singular Structures, Aachen, Germany.
Service
Editorial boards
- Journal of Machine Learning Research (JMLR), Action Editor since 2025
- Transactions on Machine Learning Research (TMLR), Action Editor since 2022
- Journal of Optimization Theory and Applications (JOTA), Area Editor since 2021
Programme committees
- Area Chair for NeurIPS, ICML, ICLR and AAAI
- Cluster Chair for Optimization for Data Science and Machine Learning, ICCOPT 2022
Workshops and community
- Co-organizer of eight editions of the NeurIPS Workshop on Optimization for Machine Learning (2019–2026)
- Co-organizer of the Federated Learning One World Seminar (2022–2024)
- Member of SIAM, ACM and ELLIS
Open positions & how to apply
I am interested in hearing from PhD candidates, postdocs, students and visiting researchers whose interests overlap with the group's research. See the dedicated page for current opportunities and application instructions.
- PhD position Fully funded, four years, on the E13 scale.
- Postdoc Funded by the group, or hosted with your own fellowship.
- HiWi Paid research work for students in Germany, before completing a master's degree.
- Internship or visit For international students and visiting researchers.
Positions & education
- since 2025Tenured Faculty, CISPASaarbrücken, Germany
- 2021–2025Tenure-Track Faculty, CISPASaarbrücken, Germany
- 2016–2021Research Scientist, EPFLHost: Martin Jaggi. Lausanne, Switzerland
- 2014–2016SNSF Postdoctoral Fellow, UCLouvainHosts: Yurii Nesterov and François Glineur. Louvain-la-Neuve, Belgium
- 2010–2014PhD in Computer Science, ETH ZurichConvex Optimization with Random Pursuit. Advisors: Bernd Gärtner and Christian L. Müller
- 2008–2010MSc in Mathematics, ETH Zurich, with distinctionGraph sparsification and applications