Philip Naumann
I am a PhD student in Machine Learning at the Berlin Institute for the Foundations of Learning and Data (BIFOLD) at Technische Universität Berlin, supervised by Prof. Klaus-Robert Müller and Prof. Grégoire Montavon. My research centers on explaining and modeling distribution shifts through the lens of optimal transport and XAI. I am interested in both foundational questions and applications in settings where model reliability and robustness matter, such as digital pathology and industrial processes.News
- Jun 2026Paper Towards Robust Foundation Models for Digital Pathology published in Nature Communications.
- May 2026New preprint: Reliable Modeling of Distribution Shifts via Displacement-Reshaped Optimal Transport on arXiv.
- Apr 2026Co-organized the Special Session on Efficient and Resilient Machine Learning for Industrial Applications at ESANN 2026.
- Jan 2026Paper Wasserstein Distances Made Explainable published in IEEE Transactions on Pattern Analysis and Machine Intelligence.
Publications
Selected Publications
Nature Communications, 17(1):5218, 2026
TL;DR Pathology foundation models often pick up on scanner and lab artifacts rather than biology. We introduce PathoROB, a benchmark of new robustness metrics spanning multiple biological classes and medical centers, showing robustness gaps across the models tested and arguing it must become a core design goal before clinical use.
@article{komen2026pathology,
title = {Towards Robust Foundation Models for Digital Pathology},
author = {Jonah K{\"o}men and Edwin D. de Jong and Julius Hense and Hannah Marienwald and Jonas Dippel and Philip Naumann and Eric Marcus and Lukas Ruff and Maximilian Alber and Jonas Teuwen and Frederick Klauschen and Klaus-Robert M{\"u}ller},
journal = {Nature Communications},
year = {2026},
volume = {17},
number = {1},
pages = {5218},
doi = {10.1038/s41467-026-73923-2}
}
arXiv preprint, 2026
TL;DR Optimal transport under standard Euclidean cost can misrepresent distribution shifts. We reshape the ground metric using a Mahalanobis distance derived from observed sample displacements, giving more reliable transport plans while staying compatible with existing OT solvers.
@misc{naumann2026reshapeOT,
title = {Reliable Modeling of Distribution Shifts via Displacement-Reshaped Optimal Transport},
author = {Philip Naumann and Jacob Kauffmann and Klaus-Robert M{\"u}ller and Gr{\'e}goire Montavon},
year = {2026},
eprint = {2605.04965},
archivePrefix = {arXiv},
primaryClass = {cs.LG},
url = {https://arxiv.org/abs/2605.04965}
}
IEEE Transactions on Pattern Analysis and Machine Intelligence, 48(6):6393-6406, 2026
TL;DR Knowing the Wasserstein distance between two distributions doesn’t tell you why it’s large or small. We propose an Explainable-AI approach that attributes the distance to data subgroups, input features, or interpretable subspaces, making dataset shifts and transport phenomena easier to understand.
@article{naumann2026wax,
title = {Wasserstein Distances Made Explainable: Insights Into Dataset Shifts and Transport Phenomena},
author = {Philip Naumann and Jacob Kauffmann and Gr{\'e}goire Montavon},
journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
year = {2026},
volume = {48},
number = {6},
pages = {6393--6406},
doi = {10.1109/TPAMI.2026.3656947}
}
Other Publications
European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (ESANN), 2026
TL;DR A survey of the efficiency and reliability challenges facing ML in industrial settings like manufacturing and logistics, covering data-centric methods, efficient training, hardware-optimized deployment, and where foundation models fit in.
@inproceedings{wissmann2026esann,
title = {Efficient and Resilient Machine Learning for Industrial Applications},
author = {Philipp Wissmann and Philip Naumann and Daniel Hein and Steffen Udluft and Marc Weber and Simon Leszek and Thomas Runkler},
booktitle = {34th European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (ESANN)},
year = {2026},
doi = {10.14428/esann/2026.ES2026-6}
}
xAI (Late-breaking Work, Demos, Doctoral Consortium), CEUR Workshop Proceedings, vol. 3793, pp. 425-432, 2024
TL;DR Optimal transport is widely used to model distribution shifts, but its own black-box nature is rarely questioned. This position paper motivates XAI-OT: explaining what drives an optimal transport solution rather than just trusting it at face value.
@inproceedings{naumann2024xai,
title = {Towards {XAI} for Optimal Transport},
author = {Philip Naumann},
booktitle = {xAI (Late-breaking Work, Demos, Doctoral Consortium)},
series = {CEUR Workshop Proceedings},
volume = {3793},
pages = {425--432},
publisher = {CEUR-WS.org},
year = {2024}
}
ECML/PKDD, Lecture Notes in Computer Science, vol. 12976, pp. 682-698, 2021
TL;DR Counterfactual explanations usually assume changes happen instantly, ignoring that real-world actions take effort and must be applied in some order. We formulate sequential counterfactual generation as multi-objective optimization and solve it with a genetic algorithm that accounts for each action’s downstream consequences.
@inproceedings{naumann2021ecml,
title = {Consequence-Aware Sequential Counterfactual Generation},
author = {Philip Naumann and Eirini Ntoutsi},
booktitle = {{ECML/PKDD}},
series = {Lecture Notes in Computer Science},
volume = {12976},
pages = {682--698},
publisher = {Springer},
year = {2021},
doi = {10.1007/978-3-030-86520-7_42}
}