When should a machine-learning system no longer be trusted?

I develop methods to recognize when machine-learning models encounter unfamiliar inputs and may behave unreliably.

Available for postdoctoral positions from early 2027. Get in touch

Figure 1. A simple example of out-of-distribution detection: a Gaussian detector accepts inputs inside the dashed contours and flags everything else. Point at the plot, or use the arrow keys, to test an input.Tap the plot to test an input.

Selected research

All 15 publications

Improving Out-of-Distribution Detection with Markov Logic Networks

Konstantin Kirchheim, Frank Ortmeier

ICML 2025

Out-of-Distribution Detection with Logical Reasoning

Konstantin Kirchheim, Tim Gonschorek, Frank Ortmeier

WACV 2024

PyTorch-OOD: A library for Out-of-Distribution Detection based on PyTorch

Konstantin Kirchheim, Marco Filax, Frank Ortmeier

CVPR Workshops (HCIS) 2022

About

I am a PhD researcher in computer science at Otto von Guericke University Magdeburg and was previously a visiting graduate student at the University of Waterloo.

For my next research role, I want to connect detection and monitoring methods with the evidence needed to deploy machine learning in safety-critical systems, particularly safe perception. I am seeking postdoctoral positions from early 2027 in trustworthy machine learning and evaluation methodology.

Lines of work

Research overview

Distribution Shift and Monitoring
Methods for recognizing inputs and operating conditions beyond those under which a model can be expected to perform reliably.
9 publications
Reasoning with Structured Knowledge
Using logical, probabilistic, and natural-language knowledge to recognize semantically inconsistent situations.
4 publications
Evaluation and Reproducibility
Software and protocols for making empirical claims about model reliability easier to compare, reproduce, and interpret.
3 publications

Software and teaching

All projects

Open-source software

`pytorch-ood`

Evaluation infrastructure for out-of-distribution detection, with unified score semantics, tested implementations, datasets, pretrained models, and benchmark interfaces.

Teaching

Introduction to Machine Learning Safety

A graduate-level course on system safety, testing, distribution shift, uncertainty, adversarial robustness, and assurance for machine-learning systems.

Talks and public engagement

Radio interview, Deutschlandfunk

Safe use of AI in railway operations

A conversation about distribution shift, reliable uncertainty estimates, and how learned components interact with established railway safety processes.