Klaus-Robert Müller

Jacob Kauffmann, Jonas Dippel, Lukas Ruff et al.

Unsupervised learning has become an essential building block of AI systems. The representations it produces, e.g. in foundation models, are critical to a wide variety of downstream applications. It is therefore important to carefully examine unsupervised models to ensure not only that they produce accurate predictions, but also that these predictions are not "right for the wrong reasons", the so-called Clever Hans (CH) effect. Using specially developed Explainable AI techniques, we show for the first time that CH effects are widespread in unsupervised learning. Our empirical findings are enriched by theoretical insights, which interestingly point to inductive biases in the unsupervised learning machine as a primary source of CH effects. Overall, our work sheds light on unexplored risks associated with practical applications of unsupervised learning and suggests ways to make unsupervised learning more robust.

Sidney Bender, Ole Delzer, Jan Herrmann et al.

Deep learning models remain vulnerable to spurious correlations, leading to so-called Clever Hans predictors that undermine robustness even in large-scale foundation and self-supervised models. Group distributional robustness methods, such as Deep Feature Reweighting (DFR) rely on explicit group labels to upweight underrepresented subgroups, but face key limitations: (1) group labels are often unavailable, (2) low within-group sample sizes hinder coverage of the subgroup distribution, and (3) performance degrades sharply when multiple spurious correlations fragment the data into even smaller groups. We propose Counterfactual Knowledge Distillation (CFKD), a framework that sidesteps these issues by generating diverse counterfactuals, enabling a human annotator to efficiently explore and correct the model's decision boundaries through a knowledge distillation step. Unlike DFR, our method not only reweights the undersampled groups, but it also enriches them with new data points. Our method does not require any confounder labels, achieves effective scaling to multiple confounders, and yields balanced generalization across groups. We demonstrate CFKD's efficacy across five datasets, spanning synthetic tasks to an industrial application, with particularly strong gains in low-data regimes with pronounced spurious correlations. Additionally, we provide an ablation study on the effect of the chosen counterfactual explainer and teacher model, highlighting their impact on robustness.

Marvin Sextro, Gabriel Dernbach, Kai Standvoss et al.

Understanding how deep learning models predict oncology patient risk can provide critical insights into disease progression, support clinical decision-making, and pave the way for trustworthy and data-driven precision medicine. Building on recent advances in the spatial modeling of the tumor microenvironment using graph neural networks, we present an explainable cell graph (xCG) approach for survival prediction. We validate our model on a public cohort of imaging mass cytometry (IMC) data for 416 cases of lung adenocarcinoma. We explain survival predictions in terms of known phenotypes on the cell level by computing risk attributions over cell graphs, for which we propose an efficient grid-based layer-wise relevance propagation (LRP) method. Our ablation studies highlight the importance of incorporating the cancer stage and model ensembling to improve the quality of risk estimates. Our xCG method, together with the IMC data, is made publicly available to support further research.

Ludwig Winkler, Klaus-Robert Müller, Huziel E. Sauceda

Molecular dynamics simulations are a cornerstone in science, allowing to investigate from the system's thermodynamics to analyse intricate molecular interactions. In general, to create extended molecular trajectories can be a computationally expensive process, for example, when running $ab-initio$ simulations. Hence, repeating such calculations to either obtain more accurate thermodynamics or to get a higher resolution in the dynamics generated by a fine-grained quantum interaction can be time- and computationally-consuming. In this work, we explore different machine learning (ML) methodologies to increase the resolution of molecular dynamics trajectories on-demand within a post-processing step. As a proof of concept, we analyse the performance of bi-directional neural networks such as neural ODEs, Hamiltonian networks, recurrent neural networks and LSTMs, as well as the uni-directional variants as a reference, for molecular dynamics simulations (here: the MD17 dataset). We have found that Bi-LSTMs are the best performing models; by utilizing the local time-symmetry of thermostated trajectories they can even learn long-range correlations and display high robustness to noisy dynamics across molecular complexity. Our models can reach accuracies of up to 10$^{-4}$ angstroms in trajectory interpolation, while faithfully reconstructing several full cycles of unseen intricate high-frequency molecular vibrations, rendering the comparison between the learned and reference trajectories indistinguishable. The results reported in this work can serve (1) as a baseline for larger systems, as well as (2) for the construction of better MD integrators.

Christopher J. Anders, Plamen Pasliev, Ann-Kathrin Dombrowski et al.

Explanation methods promise to make black-box classifiers more transparent. As a result, it is hoped that they can act as proof for a sensible, fair and trustworthy decision-making process of the algorithm and thereby increase its acceptance by the end-users. In this paper, we show both theoretically and experimentally that these hopes are presently unfounded. Specifically, we show that, for any classifier $g$, one can always construct another classifier $\tilde{g}$ which has the same behavior on the data (same train, validation, and test error) but has arbitrarily manipulated explanation maps. We derive this statement theoretically using differential geometry and demonstrate it experimentally for various explanation methods, architectures, and datasets. Motivated by our theoretical insights, we then propose a modification of existing explanation methods which makes them significantly more robust.

Thomas Schnake, Oliver Eberle, Jonas Lederer et al.

Graph Neural Networks (GNNs) are a popular approach for predicting graph structured data. As GNNs tightly entangle the input graph into the neural network structure, common explainable AI approaches are not applicable. To a large extent, GNNs have remained black-boxes for the user so far. In this paper, we show that GNNs can in fact be naturally explained using higher-order expansions, i.e. by identifying groups of edges that jointly contribute to the prediction. Practically, we find that such explanations can be extracted using a nested attribution scheme, where existing techniques such as layer-wise relevance propagation (LRP) can be applied at each step. The output is a collection of walks into the input graph that are relevant for the prediction. Our novel explanation method, which we denote by GNN-LRP, is applicable to a broad range of graph neural networks and lets us extract practically relevant insights on sentiment analysis of text data, structure-property relationships in quantum chemistry, and image classification.

David Lassner, Anne Baillot, Sergej Dogadov et al.

Alterations in historical manuscripts such as letters represent a promising field of research. On the one hand, they help understand the construction of text. On the other hand, topics that are being considered sensitive at the time of the manuscript gain coherence and contextuality when taking alterations into account, especially in the case of deletions. The analysis of alterations in manuscripts, though, is a traditionally very tedious work. In this paper, we present a machine learning-based approach to help categorize alterations in documents. In particular, we present a new probabilistic model (Alteration Latent Dirichlet Allocation, alterLDA in the following) that categorizes content-related alterations. The method proposed here is developed based on experiments carried out on the digital scholarly edition Berlin Intellectuals, for which alterLDA achieves high performance in the recognition of alterations on labelled data. On unlabelled data, applying alterLDA leads to interesting new insights into the alteration behavior of authors, editors and other manuscript contributors, as well as insights into sensitive topics in the correspondence of Berlin intellectuals around 1800. In addition to the findings based on the digital scholarly edition Berlin Intellectuals, we present a general framework for the analysis of text genesis that can be used in the context of other digital resources representing document variants. To that end, we present in detail the methodological steps that are to be followed in order to achieve such results, giving thereby a prime example of an Machine Learning application the Digital Humanities.