The Climate Informatics Group is part of German Aerospace Center’s Institute of Data Science in Jena (website there). Since 2021 another branch of the group is located at Technische Universität (TU) Berlin.
As part of the ERC Starting Grant CausalEarth another branch of the climate informatics group has been established at Technical University (TU) Berlin where I am guest professor in the computer science department since February 2021. The new branch enables lively exchange and collaborations with the world-class TU research chairs on AI and machine learning.
CausalEarth (more information here) enables several positions at TU Berlin on causal inference and discovery methods inspired by challenges of the spatio-temporal complex system Earth. You will develop theory and methods at the forefront of causal inference and AI and help climate scientists apply them to better understand and predict the climate system and climate change. The well-funded positions offer ample opportunities for research travel and more extended stays at our international partners and can also be involved in the ELLIS network among leading European AI researchers. We look for applicants with a strong background in math/stats/physics/ML. There are currently the following options (more coming soon):
- Postdoc position (official ad coming soon, but send me your application by email, see below)
We build a diverse group and especially encourage women to apply! For all positions, please send your application in ONE PDF (only CV, grades, motivation for now) to runge at tu-berlin.de.
We have a new postdoc position on “Causally Interpretable AI” that will investigate and further develop promising approaches from the field of Explainable AI, i.e. explainable machine learning models, with respect to causal interpretability. Deep Learning-based predictions have achieved exceptional performance on climate phenomena such as the El Niño/Southern Oscillation (ENSO), which far exceeded the results of physics-based predictions. The problem is, however, what the basis for this performance is and how robust it is? In other words, what “features” of the data were learned? There are a number of methods for interpreting and understanding Deep Learning techniques. These are based, for example, on Deep Taylor approaches, sensitivity maps, and relevance scores. For ENSO, such techniques indicate specific regions in tropical temperature fields. However, the question is whether these methods show features that can be interpreted causally. This would mean that a hypothetical manipulation of the temperature in these regions would also have a causal effect on ENSO. The investigation of the general connection between Explainable AI and Causal Inference is the question of this project. If you have a strong background in math/stats/physics/ML, apply here. We build a diverse group and especially encourage women to apply!
Checkout our new NeurIPS paper on high-recall causal discovery for autocorrelated time series with latent confounders. Joint work with Andreas Gerhardus. More comparisons soon on the causality benchmark platform http://causeme.net
Checkout my new UAI paper on time series causal discovery for lagged AND contemporaneous dependencies (PCMCI+). Often causal links have a shorter time lag than the resolution of the time series, which leads to contemporaneous links that cannot be directed using methods like Granger causality and also my previous method PCMCI (Science Advances paper below). PCMCI+ addresses this issue and can identify contemporaneous causal directions even between just two time series, if any of them has enough autocorrelation. Works very well also for many variables. More comparisons soon on the causality benchmark platform http://causeme.net
Identifying causal relationships and quantifying their strength from observational time series data are key problems in disciplines dealing with complex dynamical systems such as the Earth system. Data-driven causal inference in such systems is challenging since datasets are often high-dimensional and nonlinear with limited sample sizes. Here we introduce a novel method that flexibly combines linear or nonlinear conditional independence tests with a causal discovery algorithm to estimate causal networks from large-scale time series datasets. Our method outperforms state-of-the-art techniques in detection power, which opens up entirely new possibilities to discover and quantify causal networks from time series across a range of research fields. The paper is accompanied by the software package Tigramite. Also have a look at the Perspective paper and causality benchmark platform below!
Nature Communications perspective paper on causal discovery from time series
Together with co-authors, we recently published a Nature Communications perspective paper on causal discovery from time series. The perspective provides an overview of causal inference methods, identifies promising applications, and discusses methodological challenges (exemplified in Earth system sciences). I hope you find it useful for your work!
Watch a talk on the paper at INRIA Saclay here (thanks to Victor Estrade):
Causality benchmark platform www.causeme.net
In addition, we recently launched a causality benchmark platform www.causeme.net. The CauseMe platform provides ground truth benchmark datasets featuring different real challenges to assess and compare the performance of causal discovery methods. As a method developer, you can test your methods. As a method user, you can look up which causal methods are best suited for particular challenges. We need more ground truth datasets, contribute now!