Maximilian Müller

Maximilian Müller

Machine Learning Researcher

Publications

Mahalanobis++: Improving OOD Detection via Feature Normalization

Published in ICML, 2025

We show that basic assumptions underlying the Mahalanobis Distance for OOD detection are often severely violated, but can be improved by normalizing the features. The resulting method, Mahalanobis++, improves OOD detection consistently across architectures and pretraining schemes, and leads to new SOTA results.

Recommended citation: M. Müller and Matthias Hein (2025) “Mahalanobis++: Improving OOD Detection via Feature Normalization” ICML 2025 https://arxiv.org/abs/2505.18032v1

LoGex: Improved tail detection of extremely rare histopathology classes via guided diffusion

Published in MICCAI 2024 ADSMI Workshop, 2024

In this paper, we aim to detect extremely rare tail-classes as out-of-distribution data. We leverage low-rank adaption (LoRA) and diffusion guidance to generate targeted synthetic data for the detection problem and significantly improve the tail detection performance on a challenging histopathological task without losing classification accuracy on the head classes.

Recommended citation: Maximilian Müller and Matthias Hein (2024) “LoGex: Improved tail detection of extremely rare histopathology classes via guided diffusion”, MICCAI 2024 ADSMI Workshop

How to rain your ViT for OOD Detection

Published in ICLR 2024 workshop on Responsible and Reliable Foundation Models, 2024

We investigate the impact of both the pretraining and finetuning scheme on the performance of ViTs on ImageNet-scale OOD detection by analyzing a large pool of models. We find that the exact type of pretraining has a strong impact on which method works well and on OOD detection performance in general.

Recommended citation: Maximilian Müller and Matthias Hein (2024) “How to rain your ViT for OOD Detection”, ICLR 2024 R2FM Workshop https://arxiv.org/abs/2405.17447

Normalization Layers Are All That Sharpness-Aware Minimization Needs

Published in NeurIPS, 2023

We show that perturbing only the affine normalization parameters (roughly 0.1% of all parameters) in the adversarial step of SAM typically outperforms perturbing all of the parameters. This finding generalizes to different SAM variants and both BatchNorm and LayerNorm. Alternative sparse perturbation approaches do not achieve similar performance, especially not at such extreme sparsity levels.

Recommended citation: Maximilian Müller, Tiffany Vlaar, David Rolnick, and Matthias Hein (2023) “Normalization Layers Are All That Sharpness-Aware Minimization Needs”, NeurIPS 2023 https://arxiv.org/abs/2306.04226

In or Out? Fixing ImageNet Out-of-distribution detection

Published in ICML, 2023

We show that previously used datasets for evaluating OOD detection on ImageNet are severly flawed, because they contain a significant fraction of ID samples. We present NINCO, a novel hand-cleaned dataset as a solution.

Recommended citation: J. Bitterwolf*, M. Müller* and Matthias Hein (*equal contribution) (2023) “In or out? Fixing ImageNet Out-of-distribution detection” ICML 2023 https://arxiv.org/abs/2306.00826

Leveraging Stochastic Predictions of Bayesian Neural Networks for Fluid Simulations

Published in NeurIPS, 2022

We investigate uncertainty estimation and multimodality via the non-deterministic predictions of Bayesian neural networks (BNNs) in fluid simulations. To this end, we deploy BNNs in three challenging experimental test-cases of increasing complexity.

Recommended citation: Maximilian Mueller, Robin Greif, Frank Jenko, Nils Thuerey (2022) "Leveraging Stochastic Predictions of Bayesian Neural Networks for Fluid Simulations” NeurIPS 2022 workshop: Machine Learning and the Physical Sciences https://arxiv.org/abs/2205.01222

Physics-based Deep Learning

Published in world wide web, 2021

This document contains a practical and comprehensive introduction of everything related to deep learning in the context of physical simulations.

Recommended citation: Nils Thuerey, Philipp Holl, Maximilian Mueller, Patrick Schnell, Felix Trost, Kiwon Um (2021) “Physics-based Deep Learning” online Book https://physicsbaseddeeplearning. org

Fast evaluation of the current driven by electron cyclotron waves for reactor studies

Published in Physics of Plasmas, 2018

In this paper, a procedure for the evaluation of the optimum current driven by EC waves for given global parameters is proposed, which relies on a single numerical calculation of the current drive efficiency, based on the adjoint method (including momentum-conserving corrections).

Recommended citation: E Poli, M Müller, H Zohm, M Kovari (2018) “Fast evaluation of the current driven by electron cyclotron waves for reactor studies” Physics of Plasmas, vol. 25 https://pure.mpg.de/rest/items/item_3018972/component/file_3023894/content