Celestine Dünner

by Tübingen AI Center / Elia Schmid

Celestine Mendler-Dünner

I am a Principal Investigator at the ELLIS Institute Tübingen, co-affiliated with the Max Planck Institute for Intelligent Systems and the Tübingen AI Center, leading the Algorithms and Society Group. My research focus on machine learning in social context and the role of prediction in digital economies. Before joining the ELLIS Institute I was a researcher at the Max Planck Institute for Intelligent Systems and spent two years as a SNSF postdoctoral fellow at UC Berkeley hosted by Moritz Hardt. I obtained my PhD from ETH Zurich, advised by Thomas Hofmann. During my PhD I was employed at IBM Research Zurich where I co-lead the design of the IBM Snap ML library.

LinkedIn Google Scholar Twitter Email Research Projects Publications

Talks and Events

Affiliations and Awards

I am an ELLIS Scholar, a core faculty member of the Tübingen AI Center and the International Max Planck Research School for Intelligent Systems (IMPRS-IS), and an associated faculty member at the Max Planck ETH Center for Learning Systems (CLS). I am also a member of the Tübingen Cluster of Excellence on ML for Science, and a fellow of the Elisabeth-Schiemann Kolleg of the Max Planck Society. For my dissertation I was awarded the ETH Medal. For the high industrial impact of my research on system-aware algorithm design I received the IBM Research Devision Award, the IBM Eminence and Excellence Award, and the Fritz Kutter Prize. My Postdoc was supported by the SNSF Early Postdoc Mobility Fellowship by the Swiss National Science Foundation.

Advising

I am building up a research group at the ELLIS Institute in Tübingen!

Are you interested in working with me as a Postdoc? Please get in touch via email.
Prospective PhD students, please apply through one of the following PhD programs:

Resources, Media and Recordings

Podcast Interview

Featured Research Projects

My research broadly studies foundations of machine learning with a focus on social questions. I am working towards developing theoretical as well as practical tools to support safe, reliable and trustworthy machine learning with a positive impact on society. This encompasses technical challenges around interactive machine learning, optimization in dynamic environments, and resource-efficient learning, as well as interdisciplinary questions on social dynamics around algorithms, quantifying their impact on digital economies, and developing tools to support the responsible use of large language models.
LLMs and Surveys
As large language models increase in capability, researchers have started to conduct surveys of all kinds on these models---research questions rage from learning about biases and alignment of LLMs, to extracting information about human subpopulations and using them as a data source for social science research. In our work we carefully examine the validity of surveys as a method to extract population statistics from LLMs. We explain pitfalls, develop tools for systematic testing the capabilities of LLMs to emulate human outcomes, and design statistically rigorous methods to integrate them into the research pipeline.
Snap ML
Evaluating LLMs as risk scores

A Python library to systematically translate the ACS survey data into natural text prompts to benchmark LLM outputs against US population statistics. Individual prediction tasks are non-realizable and provide insights into model's ability to express natural uncertainty in human outcomes.

Performative Prediction
Predictions when deployed in societal systems can trigger actions and reactions of individuals. Thereby they can change the way the broader system behaves -- a dynamic effect that traditional machine learning fails to account for. To formalize and reason about performativity in the context of machine learning, we have developed the framework of performative prediction. It extends the classical risk minimization framework in that it allows the data distribution to depend on the deployed model. This inherently dynamic viewpoint leads to new solution concepts and optimization challenges, and it brings forth interesting connections to concepts from causality, game theory and control.
Retraining
Performative Prediction

Power in Digital Economies
Algorithmic predictions play an increasingly important role in digital economies as they mediate services, platforms, and markets at societal scale. The fact that such services can steer consumption and behavior is a central concern in modern anti-trust. I am exploring how the concept of performativity can help quantify economic power in digital economies and support digital market investigations. Intuitively, the more powerful a firm the more performative their predictions, a causal effect strength we can measure from observational and experimental data.
Powermeter
Performative power of online search

A Chrome extension that measure how content arrangements impact user click behavior through randomized experiments.

Algorithmic Collective Action
Digital platforms critically rely on data provided by individuals. Thus data offers a lever to the participants for countering power imbalances and gaining some control over the system. It can be seen a new technology specific alternative to traditional strikes. We study different learning settings and quantify the critical mass of individuals that need to be mobilized to achieve concrete goals. The effecitveness of strategies is closely related to performativity of the platform's algorithm. Research questions encompass the design of strategies, challenges of coordination, as well as connection to labor markets and collective action theory in political economy.
Powermeter
Algorithmic collective action

  • Theoretical framework. ICML 2022.
  • Application to sequential recommender systems. NeurIPS 2024.
  • A collection of documented use-cases. Github.

Scalable Learning Algorithms
When building societal-scale machine learning systems, training time and resource constraints can be a critical bottleneck for dynamic system optimization. Efficient training algorithms that are aware of distributed architectures, interconnect topologies, memory constraints and accelerator units form an important bilding block towards using available resources most efficiently. As part of my PhD research, we demonstrated that system-aware algroithms can lead to several orders of magnitude reduction in training time compared to standard system-agnostic methods. Today, the innovations of my PhD research form the backbone of the IBM Snap ML library that has been integrated with several of IBM's core AI products.
Snap ML
IBM Snap Machine Learning

Snap ML is a library that provides resource efficient and fast training of popular machine learning models on modern computing systems.
>400k downloads on PyPi
https://www.zurich.ibm.com/snapml/

Publications and Preprints

*alphabetical order
Decline Now: A combinatorial model for algorithm collective action
D.Sigg, M.Hardt and C.Mendler-Dünner
Arxiv, 2024.
Evaluating language models as risk scores
A.Cruz, M.Hardt and C.Mendler-Dünner
Advances in Neural Information Processing Systems (NeurIPS), 2024.
Questioning the Survey Responses of Large Language Models
R.Dominguez-Olmedo, M.Hardt and C.Mendler-Dünner
Advances in Neural Information Processing Systems (NeurIPS -- oral), 2024.
An engine not a camera: Measuring performative power of online search
C.Mendler-Dünner, G.Carovano and M.Hardt
Advances in Neural Information Processing Systems (NeurIPS), 2024.
Algorithmic Collective Action in Recommender Systems: Promoting Songs by Reordering Playlists
J.Baumann and C.Mendler-Dünner
European Workshop on Algorithmic Fairness (EWAF), 2024.
Advances in Neural Information Processing Systems (NeurIPS), 2024.
Causal Inference out of Control: Estimating Performativity without Tretament Randomization
G.Cheng, M.Hardt and C.Mendler-Dünner
International Conference on Machine Learning (ICML), 2024.
Performative Prediction: Past and Future
M.Hardt and C.Mendler-Dünner
Arxiv, 2023.
Collaborative Learning via Prediction Consensus
D.Fan, C.Mendler-Dünner and M.Jaggi
Advances in Neural Information Processing Systems (NeurIPS), 2023.
Algorithmic Collective Action in Machine Learning
M.Hardt*, E.Mazumdar*, C.Mendler-Dünner* and T.Zrnic*
International Conference on Machine Learning (ICML), 2023.
Anticipating Performativity by Predicting from Predictions
C.Mendler-Dünner, F.Ding and Y.Wang
Advances in Neural Information Processing Systems (NeurIPS), 2022.
Performative Power
M.Hardt*, M.Jagadeesan* and C.Mendler-Dünner*
Advances in Neural Information Processing Systems (NeurIPS), 2022.
Regret Minimization with Performative Feedback
M.Jagadeesan, T.Zrnic and C.Mendler-Dünner
International Conference on Machine Learning (ICML), 2022.
Symposium on Foundations of Responsible Computing (FORC), 2022.
Test-time Collective Prediction
C.Mendler-Dünner, W.Guo, S.Bates and M.I.Jordan
Advances in Neural Information Processing Systems (NeurIPS), 2021.
Alternative Microfoundations for Strategic Classification
M.Jagadeesan, C.Mendler-Dünner and M.Hardt
International Conference on Machine Learning (ICML), 2021.
Differentially Private Stochastic Coordinate Descent
G.Damaskinos, C.Mendler-Dünner, R.Guerraoui, N.Papandreou and T.Parnell
AAAI Conference on Artificial Intelligence (AAAI), 2021.
Stochastic Optimization for Performative Prediction
C.Mendler-Dünner*, J.C.Perdomo*, T.Zrnic* and M.Hardt
Advances in Neural Information Processing Systems (NeurIPS), 2020.
Performative Prediction
J.C.Perdomo*, T.Zrnic*, C.Mendler-Dünner and M.Hardt
International Conference on Machine Learning (ICML), 2020.
Randomized Block-Diagonal Preconditioning for Parallel Learning
C.Mendler-Dünner and A.Lucchi
International Conference on Machine Learning (ICML), 2020.
SySCD: A System-Aware Parallel Coordinate Descent Algorithm
N.Ioannou*, C.Mendler-Dünner* and T.Parnell
Advances in Neural Information Processing Systems (NeurIPS -- Spotlight), 2019.
On Linear Learning with Manycore Processors
E.Wszola, C.Mendler-Dünner, M.Jaggi and M.Püschel
IEEE International Conference on High Performance Computing (HiPC -- best paper finalist), 2019.
System-Aware Algorithms for Machine Learning
C.Mendler-Dünner
ETH Research Collection (PhD Thesis -- ETH medal), 2019.
Snap ML: A Hierarchical Framework for Machine Learning
C.Dünner*, T.Parnell*, D.Sarigiannis, N.Ioannou, A.Anghel, G.Ravi, M.Kandasamy and H.Pozidis
Advances in Neural Information Processing Systems (NeurIPS), 2018.
A Distributed Second-Order Algorithm You Can Trust
C.Dünner, M.Gargiani, A.Lucchi, A.Bian, T.Hofmann and M.Jaggi
International Conference on Machine Learning (ICML), 2018.
Addressing Interpretability and Cold-Start in Matrix Factorization for Recommender Systems
C.Dünner*, M. Vlachos*, R.Heckel, V.Vassiliaadis, T.Parnell and K.Atasu
IEEE Transactions on Knowledge and Data Engineering (TKDE), 2018.
Tera-Scale Coordinate Descent on GPUs
T.Parnell, C.Dünner, K.Atasu, M.Sifalakis and H.Pozidis
Journal of Future Generation Computer Systems (FGCS), 2018.
Efficient Use of Limited-Memory Accelerators for Linear Learning on Heterogeneous Systems
C.Dünner, T.Parnell and M.Jaggi
Advances in Neural Information Processing Systems (NIPS), 2017.
Understanding and Optimizing the Performance of Distributed Machine Learning Applications on Apache Spark
C.Dünner, T.Parnell, K.Atasu, M.Sifalakis and H.Pozidis
IEEE International Conference on Big Data (IEEE Big Data), 2017.
High-Performance Recommender System Training Using Co-Clustering on CPU/GPU Clusters
K.Atasu, T.Parnell, C.Dünner, M.Vlachos and H.Pozidis
International Conference on Parallel Processing (ICPP), 2017.
Scalable and Interpretable Product Recommendations via Overlapping Co-Clustering
R.Heckel, M.Vlachos, T.Parnell and C.Dünner
IEEE International Conference on Data Engineering (ICDE), 2017.
Primal-Dual Rates and Certificates
C.Dünner, S.Forte, M.Takac and M.Jaggi
International Conference on Machine Learning (ICML), 2016.

Peer-reviewed Workshop Contributions

Questioning the Survey Responses of Large Language Models
R.Dominguez-Olmedo, M.Hardt and C.Mendler-Dünner
Workshop on Reliable and Responsible Foundation Models (R2-FM@ICLR -- oral), 2024.
Revisiting Design Choices in Proximal Policy Optimization
C.C.-Y.Hsu, C.Mendler-Dünner and M.Hardt
Workshop on Real World Challenges in RL (RWRL@NeurIPS), 2020.
Differentially Private Stochastic Coordinate Descent
G.Damaskinos, C.Mendler-Dünner, R.Guerraoui, N.Papandreou and T.Parnell
Workshop on Privacy Preserving ML (PPML@NeurIPS), 2020.
Breadth-first, Depth-next Training of Random Forests
A.Anghel*, N.Ioannou*, T.Parnell, N.Papandreou, C.Mendler-Dünner and H.Pozidis
Workshop on Systems for ML (MLSys@NeurIPS), 2019.
Snap ML
C.Mendler-Dünner and A.Anghel
Women in Machine Learning Workshop (WiML@NeurIPS), 2018.
Sampling Acquisition Functions for Batch Bayesian Optimization
A.De Palma, C.Mendler-Dünner, T.Parnell, A.Anghel and H.Pozidis
Workshop on Bayesian Nonparametrics (BNP@NeurIPS), 2018.
Parallel training of linear models without compromising convergence
N.Ioannou, C.Mendler-Dünner, K.Kourtis, T.Parnell
Workshop on Systems for ML (MLSys@NeurIPS), 2018.
Large-Scale Stochastic Learning using GPUs
T.Parnell, C.Dünner, K.Atasu, M.Sifalakis and H.Pozidis
IEEE International Workshop on Parallel and Distributed Computing for Large Scale Machine Learning and Big Data Analytics (ParLearning), 2017.

Patents

US11803779B2 - T. Parnell, A. Anghel, N. Ioannou, N. Papandreou, C. Mendler-Dünner, D. Sarigiannis, H. Pozidis.
US11573803B2 - N. Ioannou, C. Dünner, T. Parnell.
US11562270B2 - M. Kaufmann, T. Parnell, A. Kourtis, C. Mendler-Dünner.
US11461694B2 - T. Parnell, C. Dünner, D. Sarigiannis, H. Pozidis.
US11315035B2 - T. Parnell, C. Dünner, H. Pozidis, D. Sarigiannis
US11301776B2 - C. Dünner, T. Parnell, H. Pozidis.
US11295236B2 - C. Dünner, T. Parnell, H. Pozidis.
US10147103B2 - C. Dünner, T. Parnell, H. Pozidis, V. Vasileiadis, M. Vlachos.
US10839255B2 - K. Atasu, C. Dünner, T. Mittelholzer, T. Parnell, H. Pozidis, M. Vlachos.