I’m a doctoral researcher in probabilistic machine learning jointly supervised by Niki Kilbertus (TUM) and Andreas Krause (ETH) under the ELLIS PhD Programme, and funded by Helmholtz AI. I work on sequential decision making, which includes the theory and methods of reinforcement learning, world models, and causal inference. I’m also interested in scientific and robotics applications of these methods.
Previously, I completed an MSc in mathematics at Oxford and an MSc in informatics at TUM, as well as a student exchange at UNSW Sydney. I also did two software engineering internships at Google in the US and worked as a research assistant in the group of Stephan Günnemann. My studies were supported by scholarships from Studienstiftung, Max Weber Programme, DAAD, and Lincoln College, University of Oxford.
Climate change exacerbates riverine floods, which occur with higher frequency and intensity than ever. The much-needed forecasting systems typically rely on accurate river discharge predictions. To this end, the SOTA data-driven approaches treat forecasting at spatially distributed gauge stations as isolated problems, even within the same river network. However, incorporating the known topology of the river network into the prediction model has the potential to leverage the adjacency relationship between gauges. Thus, we model river discharge for a network of gauging stations with GNNs and compare the forecasting performance achieved by different adjacency definitions. Our results show that the model fails to benefit from the river network topology information, both on the entire network and small subgraphs. The learned edge weights correlate with neither of the static definitions and exhibit no regular pattern. Furthermore, the GNNs struggle to predict sudden, narrow discharge spikes. Our work hints at a more general underlying phenomenon of neural prediction not always benefitting from graphical structure and may inspire a systematic study of the conditions under which this happens.
@inproceedings{kirschstein2024merit,series={{ICML}'24},title={The Merit of River Network Topology for Neural Flood Forecasting},booktitle={International Conference on Machine Learning (ICML)},publisher={JMLR.org},author={Kirschstein, Nikolas and Sun, Yixuan},year={2024},}
Deep Active Learning for Detection of Mercury’s Bow Shock and Magnetopause Crossings
Sahib Julka, Nikolas Kirschstein, Michael Granitzer, and 2 more authors
In European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML PKDD), 2022
I received the Faculty Award of the Computer Science and Mathematics Department of the University of Passau for my bachelor’s thesis that this paper was based on.
Accurate and timely detection of bow shock and magnetopause crossings is essential for understanding the dynamics of a planet’s magnetosphere. However, for Mercury, due to the variable nature of its magnetosphere, this remains a challenging task. Existing approaches based on geometric equations only provide average boundary shapes, and can be hard to generalise to environments with variable conditions. On the other hand, data-driven methods require large amounts of annotated data to account for variations, which can scale up the costs quickly. We propose to solve this problem with machine learning. To this end, we introduce a suitable dataset, prepared by processing raw measurements from NASA’s MESSENGER (MErcury Surface, Space ENvironment, GEochemistry, and Ranging) mission and design a five-class supervised learning problem. We perform an architectural search to find a suitable model, and report our best model, a Convolutional Recurrent Neural Network (CRNN), achieves a macro F1 score of 0.82 with accuracies of approximately 80% and 88% on the bow shock and magnetopause crossings, respectively. Further, we introduce an approach based on active learning that includes only the most informative orbits from the MESSENGER dataset measured by Shannon entropy. We observe that by employing this technique, the model is able to obtain near maximal information gain by training on just two Mercury years worth of data, which is about 10% of the entire dataset. This has the potential to significantly reduce the need for manual labeling. This work sets the ground for future machine learning endeavors in this direction and may be highly relevant to future missions such as BepiColombo, which is expected to enter orbit around Mercury in December 2025.
@inproceedings{julka2022deep,title={Deep Active Learning for Detection of Mercury's Bow Shock and Magnetopause Crossings},author={Julka, Sahib and Kirschstein, Nikolas and Granitzer, Michael and Lavrukhin, Alexander and Amerstorfer, Ute},booktitle={European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML PKDD)},year={2022},}
Deep Active Learning for Detection of Mercury’s Bow Shock and Magnetopause CrossingsIn European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML PKDD), 2022