Research Group Leader
Currently, I am a group leader for the causality group at the MPI Tuebingen.
My work focuses mainly on causal inference: we try to learn causal structures either from purely observational data or from a combination of observational and interventional data. We therefore develop both theory and methodology. Our work relates to areas like high-dimensional statistics, computational statistics or graphical models. It's an exciting research area with lots of open questions!
Before joining Tuebingen, I was a postdoc (Marie Curie fellowship) at the Seminar für Statistik, ETH Zurich (CH). During my PhD and Postdoc I have been working with Dominik Janzing and Bernhard Schölkopf at the MPI for Intelligent Systems, Tübingen (GER), and later with Peter Bühlmann and Nicolai Meinshausen at ETH Zurich. I have spent three months with Leon Bottou at Microsoft Research (WA, USA) in 2011 and two months with Martin Wainwright at UC Berkeley (CA, USA) in 2013. In 2014, I have been working with Peter Spirtes at CMU (Pittsburgh, USA) for two months. I studied Mathematics in Heidelberg (GER) and in Cambridge (UK).
- In August, I will join the statistics group in the Department of Mathematical Sciences at the University of Copenhagen as an associate professor.
- Aug 22nd - 26th, I will teach at a summer course on machine learning at the Technical University of Denmark in Copenhagen.
- May 11th - May 21st, I will give a causality lecture at the MLSS in Cádiz, Spain.
- Niklas successfully finished his master thesis (congratulations!), which resulted in this arxiv paper.
- In spring semester 2016, I am lecturing the Seminar for Statistics: Learning Blackjack at ETH Zurich.
- Our paper on invariant prediction got accepted as a discussion paper at JRSS, Series B.
Causality in 4 Steps
1. Consider the following problem: we are given data from gene A (or B) and a phenotype. Clearly, both variables are correlated. What is the best prediction for the phenotype given we are deleting gene A (or B), such that its activity becomes zero?
2. Causality matters: Intuitively, the optimal prediction should depend on the underlying causal structure:
But then, if we do not accept any form of causal notion, we cannot distinguish between these two cases and our best prediction must be: "I do not know."!
3. Causal Model: If we want to be able to describe the above situation properly, we need a so-called causal model that (1) models observational data and (2) interventional data (e.g., the distribution that arises after the gene deletion) and that (3) outputs a graph. Functional Causal Models (also called Structural Equation Models) are one class of such models, see the figure on the right. If you are interested in more details, see the script below, for example.
4. Examples of questions that are studied in this field: How can one compute intervention distributions from the graph and the observational distribution efficiently? What if some of the variables are unobserved? What are nice graphical representations? Under which assumptions can we reconstruct the causal model from the observational distribution ("causal discovery")? What if we are also given data from some of the intervention distributions? Does causal knowledge help in more "classical" tasks in machine learning and statistics?
I have written a script on causality that I am more than happy to receive feedback on. Please note that it is still missing some sections. It can be downloaded here.
Invited Talks: Conferences and Workshops
- Mar 2016: Workshop on Computationally and Statistically Efficient Inference
for Complex Large-scale Data, Oberwolfach
- Nov 2015: Workshop on "Exploring the earth system data cube", Jena
- Oct 2015: Tutorial at GCPR, Aachen
- Sep 2015: DMV - Minisymposium Statistics on Complex Structures, Hamburg
- Jul 2015: ISI - World Statistics Congress, Rio de Janeiro
- Mar 2015: Workshop on Big Data in Health Policy, Toronto
- Mar 2015: Causation from Correlation?, Die Junge Akademie, Ohlstadt
- Dec 2014: ERCIM, Working Group CMStatistics, Pisa
- Jul 2014: IMS Annual Meeting, Sydney
- Jun 2014: Workshop on Simplicity and Causal Discovery, Pittsburgh
- Dec 2013: NIPS Workshop on Causality, Lake Tahoe
- Sep 2012: Networks: Processes and Causality, Menorca
- Sep 2010: International Symposium on Quantum Thermodynamics, Stuttgart
- Oct 2009: Machine Learning approaches to statistical dep. and causality, Schloss Dagstuhl
Invited Talks: Research Visits
- Feb 2016: Department of Statistics, University of Oxford, Oxford
- Jan 2015: University of St. Andrews, St Andrews
- Jan 2015: Statslab, University of Cambridge, Cambridge
- Jan 2015: Microsoft Research, Cambridge
- Jan 2015: UCL Seminar Series, London
- Jan 2015: WIAS, Berlin
- Jul 2014: UCLA, Los Angeles
- Jul 2014: Caltech, Pasadena
- Mar 2014: University of Regensburg, Regensburg
- Jan 2014: University of Amsterdam, Amsterdam
- Jan 2014: University of Nijmegen, Nijmegen
- Nov 2013: UC Berkeley, Berkeley
- Dec 2012: IST Austria, Vienna
- Nov 2012: MPI for Dynamics and Self-Organization, Goettingen
- Jun 2012: Teleconference Causality
- Jun 2011: Seminar for Statistics, ETH Zurich, Zurich
- Jun 2010: MPI for Biogeochemistry, Jena
- Mar 2009: Teleconference Causality
- ACM Transactions on Intelligent Systems and Technology
- Annals of Statistics
- Bernoulli Journal
- IEEE Transactions of Pattern Analysis and Machine Intelligence
- IEEE Information Theory
- Journal of American Statistical Association
- Journal of Causal Infernece
- Journal of Machine Learning Research
- Statistics and Computing
- AISTATS 2015
- ICML (2012, 2013, 2014)
- ICONIP 2011
- IEEE Int. Workshop on ML for Signal Proc. (2012)
- NIPS (2011, 2015)
- COLT 2015
- UAI (2012, 2013, 2014, 2015, 2016)