Machine learning has developed into a key
technology. Its applications are everywhere, they are changing the
world we are living in, and they are influencing the way scientific
discoveries take place.
However, machine learning algorithms are never perfect - an algorithm that
always makes a correct prediction cannot exist. Rather, different
algorithms are based on different assumptions, make
different types of mistakes, and have different strengths and weaknesses.
This is where our research comes in.
Our focus is to
understand machine learning algorithms from a theoretical point of
view, understand their implicit mechanisms, and to give formal
statistical guarantees for their performance. This enables us to reveal
fundamental assumptions, biases and strenghts and weaknesses of widely used
machine learning algorithms.
Explainable machine learning
The outcome of many machine learning algorithms is so complex that
their behaviour is hard to understand, even for experts. The field of
explainable machine learning tries to solve this problem by
constructing algorithms that ``explain'' the model's behaviour. In our
research group we investigate when or whether such explanation
algorithms are reliable or not; many of our findings are negative,
indicating that explanation techniques cannot achieve what one hopes
for:
Informative Post-Hoc Explanations
Only Exist for Simple Functions. Eric Günther, Balázs Szabados, Robi Bhattacharjee, Sebastian Bordt, Ulrike von Luxburg, 2025.
link
Performative Validity of Recourse Explanations. Gunnar König, Hidde Fokkema, Timo Freiesleben, Celestine Mendler-Dünner, Ulrike von Luxburg. Arxiv, 2025.
link. NeurIPS 2025.
How to safely discard features based on aggregate SHAP values.
Robi Bhattacharjee, Karolin Frohnapfel, Ulrike von Luxburg. COLT, 2025.
link to arxiv
Sebastian Bordt, Eric Raidl, Ulrike von Luxburg:
Rethinking Explainable Machine Learning as Applied Statistics. ICML, 2025.
link to arxiv
Disentangling Interactions and Dependencies in Feature
Attribution. Gunnar König, Eric Günther, Ulrike von Luxburg, AISTATS, 2025.
link to arxiv
Robi Bhattacharjee, Ulrike von Luxburg: Auditing Local
Explanations is Hard. NeurIPS, 2024. link to arxiv
From Shapley Values to Generalized Additive Models and back. 2022.
link to arxiv
Explaining the Explainer: A First Theoretical Analysis of LIME, AISTATS 2020.
pdf
Can technical systems meet legal requirements?
We regularly collaborate with legal scholars to investigate whether
legal requirements, say from the European AI Act or the GDPR, can be
met by technical systems:
Eric Günther, Balázs Szabados, Gunnar König, Kristof Meding, Sebastian Bordt, Ulrike von Luxburg:
We Need Explanation Cards to Connect Explanation Algorithms to the Real World. 2026.
link. AIES 2026.
Karolin Frohnapfel, Mara Seyfert, Sebastian Bordt, Ulrike von
Luxburg, and Kristof Meding: Using predictive multiplicity to measure individual
performance within the AI Act. FAccT, 2026.
arxiv version
Post-Hoc Explanations Fail to Achieve their Purpose in Adversarial Contexts, FACCT 2022.
link to arxiv
Deep learning and language models
Sebastian Bordt, Martin Pawelczyk: Train Once, Answer All: Many Pretraining Experiments for the Cost of One. ICLR 2026.
Gunnar König, Martin Pawelczyk, Ulrike von Luxburg, Sebastian Bordt,
“Validity Threats for Foundation Model Research“. Arxiv, 2026
On the Surprising Effectiveness of Large Learning Rates under Standard Width Scaling.
Moritz Haas, Sebastian Bord, Ulrike von Luxburg, Leena Chennuru Vankadara, 2025.
link NeurIPS 2025.
How much can we forget about Data Contamination? Sebastian Bordt,
Suraj Srinivas, Valentyn Boreiko, Ulrike von Luxburg. ICML, 2025.
link to arxiv
When machine learning algorithms provably fail
Most machine learning researchers work on deriving new or better
machine learning algorithms. In their analysis, the focus is to
demonstrate that their algorithms work well in many cases. We find it
intriguing to reveil under which circumstances popular machine
learning algorithms provably fail. Some example publications in this
direction are:
Understanding implicit assumptions and hidden biases
Machine learning algorithms always come with hidden assumptions and
biases. However, these are rarely stated explicitly; they are implicit
consequences of the way the algorithm has been built. Part of our
work is to make these biases explicit, formalize them in the language
of mathematics, and to reveal the implicit inductive principles that
are invoked by particular algorithms. Some examples for this line
of work are the following papers:
NetGAN without GAN: From Random Walks to Low-Rank Approximations, 2020
pdf
Explaining the Explainer: A First Theoretical Analysis of LIME, 2020
pdf
How the result of graph clustering methods depends on the construction of the graph, 2013.
pdf
Phase transition in the family of p-resistances, 2011
pdf
Formal mathematical guarantees
Machine learning algorithms are never perfect - yet some of them can
be proven to survive some basic suitability tests, while others do
not. For example, one might be able to prove that an algorithm is
statistically consistent: when the algorithm gets to see an infinite
amount of data, its prediction performance is going to converge to the
best possible one. Part of our work consists in giving such formal
guarantees. Example papers are:
Foundations of Comparison-Based Hierarchical Clustering, 2019
pdf
