Lamarr Scientific Forum

Sep 3, 2025, 8:30 AM → Sep 4, 2025, 6:00 PM Europe/Berlin

Hörsaalgebäude, Campus Poppelsdorf, Universität Bonn

Welcome to the Lamarr Scientific Forum!

The Lamarr Scientific Forum is a two-day event that brings together the entire Lamarr community – from students to principal investigators – as well as a few close friends and partners.

The Scientific Forum will offer an important opportunity to connect across Lamarr’s locations and research areas, to reflect on current work, and to explore ideas for the future. It also provides space to recognize the contributions of our members and to strengthen our shared vision.

The meeting will take place at the University of Bonn. Accommodation information is available under the Accommodation tab to the left. Information about the program is updated regularly and can be found under the Timetable tab to the left.

Please make sure to register using the Registration tab to the left. The last day to register is Monday, August 18.

For any questions, feel free to contact Brendan Balcerak Jackson at bbalcera@cs.uni-bonn.de, Zoé Sánchez at zsanchez@uni-bonn.de or Ann-Kathrin Oster at ann-kathrin.oster@tu-dortmund.de. 

 

We look forward to welcoming all Lamarr members to Bonn in September!

Wed, September 3

8:30 AM Arrival & Registration

Welcome and Overview: "Lamarr Today and Tomorrow"

We begin the Lamarr Scientific Forum with a welcoming address by the Lamarr Directorate. In their plenary presentation, the directors provide an overview of the current state, activities and successes of the Lamarr Institute.

Lamarr Area Updates: Session 1

Working and growing together: On each day, 5 of the 10 Lamarr areas will provide the attendees with an update on its respective strategic focus, breakthrough results and opportunities for cooperation.

1 Area Presentation: Hybrid ML

This presentation will introduce the Hybrid ML research area, which aims to integrate deep learning with structured knowledge from mathematics and the natural and social sciences.

Hybrid ML is guided by the observation that both mathematics and the sciences can be seen as generators of compressed pattern representations. We will explain that the central goal of Hybrid ML is to align the representations inferred by deep learning models with those derived from theoretical frameworks. This goal is motivated by a fundamental hypothesis: that such aligned representations are not only easier to learn, but also more likely to contain optimal solutions (i.e., robust, interpretable, and efficient representations) tailored to the target objective. We will illustrate how this alignment can be implemented in practice, drawing on recent breakthrough results from the Hybrid ML area, and conclude with an overview of the collaboration channels available to Lamarr researchers.

Following this introduction, Paul Roetzer will give a spotlight presentation on their recent ICCV paper, “Fast Globally Optimal and Geometrically Consistent 3D Shape Matching”, in which they align neural representations with novel surface representations - in terms of cyclic paths - to obtain global, geometrically consistent matchings between three-dimensional shapes.

Speakers: Dr Ramsés Sánchez (Lamarr institute, University of Bonn), Paul Roetzer

2 Area Presentation: Planning & Logistics

The Planning & Logistics area within the Lamarr Institute focuses on transferring AI research into real-world logistics applications. Logistics offers a rich field for AI with significant impact on both society and sustainability. Key topics include scalable multi-criteria optimization for efficiency and environmental performance (such as route planning, fleet management, and navigation), expert driven process modeling (for example, in network and warehouse design), AI based digitalization (including computer vision and tracking), and intelligent automation of physical and digital processes. The presentation outlines the area’s main objective of bridging AI research and logistics practice, and highlights opportunities for collaboration, including available datasets and open challenges in logistics. Current research activities to be briefly introduced include intralogistics scene generation and sensor-based activity recognition. The session will also feature a spotlight by Nilah Nair (FLW, TU Dortmund) on a novel human motion analysis approach that extracts and transfers individual motion characteristics, enabling new applications in robotics simulation and industrial data anonymization.

Speakers: Nilah Nair, Alice Kirchheim, Anike Murrenhoff

3 Area Presentation: Resource-Aware Machine Learning

Area presentation: "Resource-Aware Machine Learning at Lamarr: A Guided Tour"

In this talk, we offer a guided overview of the resource-aware machine learning (RAML) research taking place at the Lamarr Institute. RAML aims to make machine learning systems not only accurate, but also efficient in terms of energy, latency, and computational resources. We highlight ongoing efforts within the institute and introduce some of the key researchers. Short spotlight presentations from three ongoing PhDs are presented to explore opportunities for potential collaboration. The session concludes with a brief outlook on upcoming challenges and how the Lamarr Institute aims to address them.

Break-through result: "Towards Anytime Models: A quick overview of recent results in Lamarr"

Anytime models offer flexible inference under resource constraints by producing usable predictions even when computation is interrupted. This talk outlines recent progress at the Lamarr Institute in developing such models across diverse applications. Starting from classical ideas like early exits, routing, and input-dependent computation, we show how these techniques can be extended to build full-fledged anytime systems. We highlight recent results in collaboration with the Area of Industry & Production and in the realm of image recognition, while closing the talk with a roadmap toward integrated energy-aware evaluation.

Speakers: Jian-Jia Chen, Sebastian Buschjäger (Lamarr Institute for ML and AI, TU Dortmund)

4 Area Presentation: Physics

The interdisciplinary research area of physics at the Lamarr Institute leverages advanced mathematical and machine learning methods to deepen our understanding of nature. By combining simulation-based approaches with sophisticated data analysis techniques, this area addresses fundamental questions across diverse physics domains. This presentation will introduce the research area, its objectives, and collaboration opportunities within Lamarr.

A key strength of this area is access to massive high-quality datasets and use cases from experiments spanning radio to gamma-ray astronomy, as well as astroparticle and particle physics. These datasets pose no privacy concerns and are complemented by highly precise simulations of the same experiments that provide reliable ground truth for training and testing machine learning models. Experiments such as LHCb at CERN (particle physics), the Einstein Telescope (gravitational waves), the IceCube Neutrino Observatory, the Cherenkov Telescope Array (gamma rays), and the Square Kilometer Array (radio waves) offer rich use cases for applying machine learning methods to real-world problems.

We will showcase examples of these use cases while highlighting ongoing collaborations within Lamarr and outlining new opportunities for partnership. Additionally, we will provide an overview of past successes—such as the deep-learning-based detection of neutrinos from the galactic disk or contributions to recently founded NRW Excellence Clusters with Lamarr researchers.

Spotlight Talk: Unfolding the Charm of Atmospheric Muons

The session will conclude with a spotlight presentation by Pascal Gutjahr on an ongoing research project in the IceCube subgroup in the Lamarr physics area. This collaborative work by L. Witthaus and P. Gutjahr builds on the same deep learning approach that allowed to detect neutrinos from the galactic disc. In this study, stopping and through-going muons in the IceCube detector are utilized to investigate the dominantly pion and kaon-induced atmospheric muon spectrum over a wide energy range up to the highest energies, leading towards the measurement of heavy meson decays.

Speakers: Prof. Wolfgang Rhode (TU Dortmund), Pascal Gutjahr (TU Dortmund University)

area_presentation_physics.pptx

spotlight_talk_physics.pptx

5 Area Presentation: Trustworthy AI

Within the Lamarr Institute, the topic of trustworthy AI is being explored across diverse application contexts and scientific disciplines. Lamarr researchers focus on areas such as developing effective certification and verification procedures for AI systems, ensuring explainability and robustness, as well as advancing trustworthy AI in domains like physics, life sciences, engineering, and other scientific fields. This work is complemented by broader legal, philosophical and ethical considerations related to trustworthiness of AI. Rather than attempting to cover all ongoing research within this multifaceted and highly interdisciplinary field, we will highlight two key contributions from Lamarr’s research on trustworthy AI. Both focus on the societal relevance of ensuring and implementing trustworthy AI. One spotlight talk will be given by Tim Katzke from Emmanuel Müller's research group. He will present work on "Trustworthy Machine Learning by Design." The other spotlight talk is by Rebekka Görge from Maximilian Poretschkin's group. She will discuss the trustworthiness of LLMs, focusing particularly on bias and copyright.

Speakers: Prof. Jakob Rehof (TU Dortmund), Rebekka Görge, Tim Katzke

12:45 PM Lunch

Lunch for all Lamarr members is served in the foyer of the ground floor. We encourage you to use the opportunity to network with fellow Lamarr staff.

Poster Session: Session 1

Learning and growing together: The poster session offers Lamarr members with a platform to showcase their research projects. Lamarr members across all research areas, working groups and levels of hierarchy and experience are asked to join the open exchange and discuss the presented work.

How to protect your presentation Materials.pdf

6 Adversarial Perturbations Improve Generalization of Confidence Prediction in Medical Image Segmentation

Trustworthy methods for medical image segmentation should come with a reliable mechanism to estimate the quality of their results. Training a separate component for confidence prediction is relatively fast, and can easily be adapted to different quality metrics. However, the resulting estimates are usually not sufficiently reliable under domain shifts, for example when images are taken with different devices. We introduce a novel adversarial strategy for training confidence predictors for the widely used U-Net architecture that greatly improves such generalization. It is based on creating adversarial image perturbations, aimed at substantially decreasing segmentation quality, via the gradients of the confidence predictor, leading to images outside of the original training distribution. We observe that these perturbations initially have little effect on segmentation quality. However, including them in the training gradually improves the confidence predictor's understanding of what actually affects segmentation quality when moving outside of the training distribution. On two medical image segmentation tasks, we demonstrate that this strategy significantly improves direct quality estimates and outperforms a more computationally intensive state-of-the-art method - which only produces relative, rather than absolute, scores - on volumetric and surface Dice for out-of-distribution images.

Speaker: Jonathan Lennartz

poster_jl.pdf

7 Cluster-based prediction of chatter vibrations in milling operations

Accurate predictions of process characteristics in milling, such as tool vibrations, allow for identifying and avoiding unstable cutting conditions that can lead to excessive tool wear, surface defects or tool breakage. Therefore, considering vibrations during process design is essential to ensure dimensional accuracy, surface integrity and the longevity of cutting tools. Common methods, including analytical and simulation-based approaches, often require simplifications and assumptions that limit the accuracy, applicability and scalability in real-world scenarios. In contrast, data-based modeling strategies provide predictions of specific process characteristics with high generalization capabilities. Advancements in engineering technology allow the machining of complex workpiece designs in a single setup, resulting in highly complex milling paths with changing engagement conditions. Hence, for such processes, the demands on the required training data are increased and the resulting prediction accuracy of data-based models can vary significantly for different segments of the milling path. To mitigate this issue, unsupervised learning strategies capable of identifying similar patterns in datasets can be used for manufacturing applications to group machining operations into elementary process sections. The research presented in this paper expands on this idea by training data-based models for each identified section to predict relevant process characteristics, such as tool vibrations, using acquired measurement data and tool path information as features. Thus, a milling process with highly variant engagement conditions was represented by multiple cluster-specific models, leading to an increased prediction accuracy of process characteristics compared to using only a single model.

Speaker: Florian Wöste

8 Completing the Puzzle of Unseen Perception in Grasping and Navigation

We address two critical capabilities required for autonomous robots operating in indoor environments, both centered around robust perception of unseen objects. This generalization can support various applications, but here we focus on mobile robotics.

The first focus is on robotic grasping, where 6D pose estimation is needed for successful manipulation. While 6D tracking is now reliable, the main bottleneck lies in 2D segmentation of unseen objects. Grasp success remains limited by the ability to accurately segment novel instances. Our goal is to bring 2D segmentation to practical levels for reliable, generalizable grasping.

The second focus is on high-speed navigation, where mobile robots must avoid dynamic, previously unseen obstacles in real time. To enable this, we develop a model for 3D bounding box detection of moving objects using event cameras. These sensors offer low-latency, high-temporal-resolution input essential for detecting fast motion and enabling rapid trajectory adjustments.

Together, these efforts target key missing pieces in current perception pipelines for handling unseen objects. We collected two datasets, MR6D and MTevent, which we use to benchmark models and improve performance in both segmentation and 3D detection tasks. By advancing segmentation for grasping and enabling fast 3D detection for navigation, this work contributes toward more adaptable and capable robotic systems.

---

Figure 1: Example from the MR6D dataset showing an image from the O3dyn robot camera with projected 6D annotation of a pallet. No pipeline can robustly predict 6D poses for several challenging cases, such as pallets, included in the dataset.

Figure 2: Example from the MTevent dataset showing 3D bounding box annotations for a forklift captured using our stereo-event + RGB camera system. Most existing work on moving object detection with event cameras focuses on simple scenes and 2D detection. No current method can handle realistic scenarios as shown in this image.

Speaker: Anas Gouda (TU Dortmund)

MR6D.png

MTevent.jpg

Unseen_Puzzle.pdf

9 Context-Based Meta Reinforcement Learning for Robust and Adaptable Peg-in-Hole Assembly Tasks

Autonomous assembly is an essential capability for industrial and service robots, with Peg-in-Hole (PiH) insertion being one of the core tasks. However, PiH assembly in unknown environments is still challenging due to uncertainty in task parameters, such as the hole position and orientation, resulting from sensor noise. Although context-based meta reinforcement learning (RL) methods have been previously presented to adapt to unknown task parameters in PiH assembly tasks, the performance depends on a sample-inefficient procedure or human demonstrations. Thus, to enhance the applicability of meta RL in real-world PiH assembly tasks, we propose to train the agent to use information from the robot’s forward kinematics and an uncalibrated camera. Furthermore, we improve the performance by efficiently adapting the meta-trained agent to use data from force/torque sensor. Finally, we propose an adaptation procedure for out-of-distribution tasks whose parameters are different from the training tasks. Experiments on simulated and real robots prove that our modifications enhance the sample efficiency during meta training, real-world adaptation performance, and generalization of the context-based meta RL agent in PiH assembly tasks compared to previous approaches.

Speaker: Ahmed Shokry

poster_lamarr_science_forum.pdf

10 Deep Learning for Real-time Classification of Astronomical Radio Signals: Evaluating the Applicability of Synthetic Data for Training

This work explores the applicability of synthetic data for training deep learning models aimed at real-time classification of astronomical radio signals. Building on previous research where lightweight convolutional neural networks (CNNs) using DM-time representations showed promising performance in detecting transient signals, we now turn to the question of whether synthetic datasets can serve as a reliable substitute for real observational data during training.

Synthetic data offers the advantage of full control over signal characteristics, allowing us to simulate a wide range of astrophysical phenomena and noise conditions. In this study, we generate a set of synthetic DM-time images designed to replicate realistic signal dispersion, varying signal-to-noise ratios (SNRs), receiver noise patterns, and other instrumental effects. The synthetic dataset is informed by parameters derived from well-known sources such as the Crab Pulsar and is intended to reflect the diversity and complexity of real-world radio observations.

We train minimalist CNN architectures—optimized for low-latency and low-resource environments—exclusively on synthetic data. These models will be evaluated on real pulsar observations to assess generalization capabilities across key performance metrics, including classification accuracy, precision, recall, and sensitivity to weak signals.

By comparing synthetic-trained models against baselines trained on real data, we aim to quantify the effectiveness and limitations of using simulated data in machine learning pipelines for radio astronomy. This work seeks to clarify the role synthetic data can play in accelerating model development, especially in scenarios where annotated real datasets are scarce or difficult to obtain.

Speaker: Mr Andrei Kazantsev (Max Planck Institute for Radio Astronomy)

poster_kazantsev.pdf

11 Fast, Registration-Free, and Lesion- Robust Tractography Parcellation using a Transformer Model

Tractography enables the reconstruction of white matter pathways from diffusion MRI and is a key tool for studying brain connectivity in both research and clinical contexts. Within the overall tractography pipeline, the parcellation step assigns individual streamlines to specific anatomical bundles, or discards them as false positive detections. We introduce PETParc (Parallel Efficient Tractography Parcellation), the first transformer-based architecture for this task. We demonstrate that treating individual streamlines as tokens, and letting them exchange information via the self-attention mechanism provides state-of-the-art results for registration-free parcellation, while being two orders of magnitude faster than TractCloud, the previous state of the art. Operating directly in subject space, PETParc avoids costly registration, generalizes well to unseen healthy subjects, and even outperforms TractCloud in pathological cases. In particular, we test robustness on subjects post-hemispherotomy, where PETParc reconstructs more complete and anatomically plausible bundles. Therefore, our approach offers a scalable and robust solution for high-throughput tractography analysis.

Speaker: Justus Bisten (b-it and Institute for Computer Science II & Department for Neuroradiology, University Hospital Bonn)

lamarrposter_jb_vb_final.pdf

Method.png

12 Flatness is Necessary, Neural Collapse is Not: Rethinking Generalization via Grokking

Robust generalization is a foundational requirement for trustworthy artificial intelligence (AI), underpinning the reliability, stability, and fairness of deployed systems. Two geometric phenomena are frequently correlated with generalization: neural collapse, where internal class representations converge to a maximally simple and symmetric structure, and flatness of the loss landscape, where the model settles into a wide, flat minimum, suggesting resilience to perturbations. Identifying which, if either, causally drives generalization is critical for engineering reliable AI systems.

To disentangle causation from correlation, we leverage grokking, a training regime where generalization is delayed, creating a temporal window for causal analysis. Our experimental interventions reveal a clear asymmetry: while both neural collapse and relative flatness emerge near the onset of generalization, only flatness consistently predicts it. Models regularized away from flat solutions exhibit delayed generalization, resembling grokking even in architectures and datasets where it does not typically occur. In contrast, promoting or suppressing neural collapse has no significant effect on a model’s ability to generalize. Furthermore, we show theoretically that neural collapse implies relative flatness under classical assumptions, explaining their empirical co-occurrence.

These findings reposition relative flatness as a potentially necessary and more fundamental condition for generalization. This insight has important implications for both the scientific understanding of generalization and practical methods for trustworthy AI: research and development should prioritize techniques that promote flat minima to improve reliability, robustness, and fairness. By focusing on the geometry of stability rather than structural symmetry, this work offers an actionable pathway for building AI systems that generalize better, and are, therefore, more trustworthy.

Speaker: Ting Han

TingHan_Lamarr2025_Bonn.pdf

13 GRaMPa: Subword Regularisation by Skewing Uniform Segmentation Distributions with an Efficient Path-counting Markov Model

Stochastically sampling word segmentations from a subword tokeniser, also called subword regularisation, is a known way to increase robustness of language models to out-of-distribution inputs, such as text containing spelling errors. Recent work has observed that usual augmentations that make popular deterministic subword tokenisers stochastic still cause only a handful of all possible segmentations to be sampled. It has been proposed to uniformly sample across these instead, through rejection sampling of paths in an unweighted segmentation graph. In this paper, we argue that uniformly random segmentation in turn skews the distributions of certain segmentational properties (e.g. token lengths and amount of tokens produced) away from uniformity, which still ends up hiding meaningfully diverse tokenisations. We propose an alternative uniform sampler using the same segmentation graph, but weighted by counting the paths through it. Our sampling algorithm, GRaMPa, provides hyperparameters allowing sampled tokenisations to skew towards fewer, longer tokens. Furthermore, GRaMPa is single-pass, guaranteeing significantly better computational complexity than previous approaches relying on rejection sampling. We show experimentally that language models trained with GRaMPa outperform existing regularising tokenisers in a data-scarce setting on token-level tasks such as dependency parsing, especially with spelling errors present.

Speaker: David Kaczér (University of Bonn)

GRaMPa_poster-1-1.pdf

14 Hybrid Quantum-Classical Multi-Agent Pathfinding

Multi-Agent Path Finding (MAPF) focuses on determining conflict-free paths for multiple agents navigating through a shared space to reach specified goal locations. This problem becomes computationally challenging, particularly when handling large numbers of agents, as frequently encountered in practical applications like coordinating autonomous vehicles. Quantum Computing (QC) is a promising candidate in overcoming such limits. However, current quantum hardware is still in its infancy and thus limited in terms of computing power and error robustness. In this work, we present the first optimal hybrid quantum-classical MAPF algorithms which are based on branch-and-cut-and-price. QC is integrated by iteratively solving QUBO problems, based on conflict graphs. Experiments on actual quantum hardware and results on benchmark data suggest that our approach dominates previous QUBO formulations and state-of-the-art MAPF solvers.

Speaker: Loong Kuan Lee (Fraunhofer IAIS)

15 Interaction-Aware and Domain-Invariant Representation Learning for Inclusive Flavour Tagging

Measurements of neutral, oscillating mesons are a gateway to quantum mechanics and give access to the fundamental interactions of elementary particles. For example, precise measurements of violation in neutral mesons can be taken in order to test the Standard Model of particle physics. These measurements require knowledge of the -meson flavour at the time of its production, which cannot be inferred from its observed decay products. Therefore, multiple LHC experiments employ machine learning-based algorithms, so-called flavour taggers, to exploit particles that are produced in the proton-proton interaction and are associated with the signal meson to predict the initial flavour. A state-of-the-art approach to flavour tagging is the inclusive evaluation of all reconstructed tracks from the proton-proton interaction using a Deep Set neural network.

Flavour taggers are desired to achieve optimal performance for data recorded from proton-proton interactions while being trained with a labelled data sample, i.e., with Monte Carlo simulations. However, the limited knowledge of underlying processes introduces inherent differences between simulation and recorded data. Existing flavour taggers neither model these differences nor do they model interactions between tracks explicitly, being at danger of overfitting to simulations, of not providing optimal performance for physics analyses, and of requiring a careful calibration on data.

We present an inclusive flavour tagger that builds on set transformers (to model particle interactions via set attention) and on domain-adversarial training (to mitigate differences between data sources). These foundations allow the tagger to learn intermediate data representations that are both interaction-aware and domain-invariant, i.e., they capture the interactions between tracks and do not allow for an overfitting to simulations. In our benchmark, we increase the statistical power of flavour-tagged samples by 10% with respect to the usage of deep sets, thus demonstrating the value of interaction-aware and domain-invariant representation learning.

Speaker: Quentin Führing

InclFT.pdf

16 Load Balancing Neurons: Controlling Firing Rates Improves Plasticity in Continual Learning

Neural networks often suffer from plasticity loss, which limits their ability to adapt to evolving data distributions in continual learning settings. This results in degraded performance, poor generalization, and inefficient use of model capacity. While recent methods mitigate this by resetting underutilized neurons based on utility scores, the underlying mechanisms remain poorly understood. In this work, we propose the firing rate as a simple, activation-independent metric to diagnose neuron inactivity and dysfunction, such as dead or overly linearized ReLUs. Building on this insight, we introduce a load balancing mechanism that dynamically adjusts neuron activation thresholds to maintain healthy firing rates. We further show that architectural techniques such as non-affine normalization and L2 regularization implicitly promote balanced activity and improve plasticity. Across two continual learning benchmarks, our methods lead to substantial improvements in test accuracy, surpassing both continual backpropagation and reset-based baselines.

Speaker: Jan Robine (TU Dortmund)

Load-Balancing-Neurons.pdf

17 Long-Horizon Flare Forecasting in Blazars Time Series

Forecasting astrophysical flares in blazars presents a unique challenge due to their irregular temporal dynamics and strong variability. While deep neural networks have shown promise for modeling such complex time series, their predictions often lack alignment with established physical knowledge, limiting trust and interpretability. In this work, we propose a domain-informed deep learning regularization framework that explicitly integrates astrophysical priors into the training process. Specifically, we incorporate two complementary regularization terms based on Grad-CAM explanations of model attention. First, an attention alignment loss encourages the model to focus on flux peaks, reflecting domain knowledge that flare events correspond to sharp increases in photon flux. Second, a temporal smoothness loss enforces consistency in attribution across adjacent time steps, capturing the physical duration of flare events. The combined objective function guides the model not only to minimize prediction error but also to generate explanations consistent with astrophysical priors. This approach improves both predictive accuracy and interpretability, paving the way toward more trustworthy machine learning models for flare forecasting in high-energy astrophysics.

Speaker: Amal Saadallah (Lamarr Institute-TU Dortmund)

POSTER-Physics-amal-saadallah.pdf

18 Multi-Hop Reasoning for Question Answering with Hyperbolic Representations

Hyperbolic representations are effective in modeling knowledge graph data which is prevalently used to facilitate multi-hop reasoning. However, a rigorous and detailed comparison of the two spaces for this task is lacking. In this paper, through a simple integration of hyperbolic representations with an encoder-decoder model, we perform a controlled and comprehensive set of experiments to compare the capacity of hyperbolic space versus Euclidean space in multi-hop reasoning. Our results show that the former consistently outperforms the latter across a diverse set of datasets. In addition, through an ablation study, we show that a learnable curvature initialized with the delta hyperbolicity of the utilized data yields superior results to random initializations. Furthermore, our findings suggest that hyperbolic representations can be significantly more advantageous when the datasets exhibit a more hierarchical structure.

Speaker: Akbar Karimi (University of Bonn)

welz25-acl-hyperbolic-QA-paper-poster.pptx.pdf

19 Neurodivergence in AI Research: Traits, Stress, and Well-being

The AI research ecosystem is a demanding, high-pressure environment that profoundly shapes the future of technology. Its effectiveness and sustainability depend not only on technical innovation but also on the people who sustain its progress. Investigating the psychosocial factors that link individual traits to work experiences and mental health is therefore essential for enabling sustainable, human-centered AI development.

This study investigates these pathways, particularly in 408 AI professionals, contrasting them with highly educated autistic adults (N=155) and academics in other research fields (N=72). Participants were surveyed online using standardized measures to assess autistic traits, ADHD symptoms, and work-related perceptions. Path analysis was conducted separately for each group to examine the structural relationships among variables.

The results indicate that 20.8% of AI professionals reported high rates of autistic traits and 41.2% reported ADHD symptoms, both substantially higher than prevalence estimates in the general population. The analysis shows that high stress levels directly led to increased depression and anxiety among AI professionals.
We identified specific connections: difficulties with social communication (SCI) and challenges like behavioral rigidity or sensory sensitivities (BRSR) significantly drove up this stress. Independently, low social confidence (SCRB) emerged as a strong predictor of depression symptoms. Notably, higher self-reported effort was linked to lower depression levels which demonstrates a paradoxical effect of effort. While ADHD symptoms played a role, their impact on distress was mostly indirect when we specific autistic trait facets were also considered.

The findings highlight that the AI research system relies heavily on neurodivergent individuals, whose specific trait facets can contribute to stress and impact well-being. This demands targeted neuroinclusion strategies that go beyond generic wellness programs to address trait-specific challenges. By understanding and supporting the unique needs and strengths of neurodivergent professionals, we can cultivate more resilient and truly human-centered AI research environments.

Speaker: Nicolo' Brandizzi (Fraunhofer IAIS)

Lamarr_ai_autism_presentation.pdf

20 On the Limitations of Language-Targeted Pruning for Multilingual LLMs

Current pruning methods for large language models (LLMs) achieve high compression post-training while preserving performance. However, most existing work focuses on calibration using English data, despite the multilingual nature of modern LLMs and their widespread use in non-English languages. This poster presents the first study on how the calibration language impacts pruning for multilingual LLMs in monolingual applications. The analysis spans multiple languages, tasks, models, and pruning techniques with further examination of latent subspaces, pruning masks, and individual neurons. The results reveal a critical limitation: While target-language calibration preserves general language modeling capabilities (perplexity) it does not reliably improve downstream task performance involving reasoning or knowledge retrieval. This is because pruning preserves language-specific features influenced by calibration while uniformly impairing the language-agnostic representations associated with higher-level capabilities.

Speaker: Simon Kurz

lamarr-poster-On the Limitations of Language-Targeted Pruning in Multilingual LLMs.pdf

21 Property Testing of Curve Similarity

We propose a sublinear algorithm for probabilistic testing of the discrete Fréchet distance - a standard similarity measure for curves.
We assume the algorithm is given access to the input curves via a query oracle that returns the set of vertices of the curve that lie within a radius $\delta$ of a specified vertex of the other curve.
The goal is to use a small number of queries to determine with constant probability whether the two curves have discrete Fréchet distance at most $\delta$
or they are ''$\varepsilon$-far'' (for $0 < \varepsilon < 2$) from being similar, i.e., an $\varepsilon$-fraction of the curves must be ignored for them to become similar.
We present an algorithm that is sublinear under two assumptions (i) that the curves are $\kappa$-straight, meaning they are $\kappa$-approximate shortest paths in the ambient metric space, for some $\kappa \ll n$, and (ii) that they have edge lengths within a constant factor of $\delta$ (we later relax this toward a weaker uniform sampling condition). The algorithm uses $O(\frac{\kappa}{\varepsilon} \log\frac{\kappa}{\varepsilon})$ queries and it is given the value of $\kappa$ in advance.
Our algorithm works in a matrix representation of the input and may be of independent interest to matrix testing.

Speaker: Marena Richter

PropertyTesting-Poster-Lamarr.pdf

22 Rule vs. SHAP: Complementary Tools for Understanding and Verifying ML Models

Traditional interpretability techniques such as rule-based models and feature attribution methods, each offer complementary strengths, however are often applied in isolation. Rule-based approaches are intuitive and logically structured, making them easy to understand, but they often struggle to scale effectively. On the other hand, feature attribution techniques like SHAP are well-suited to handling complex models and large datasets but can fall short in terms of interpretability and alignment with human reasoning. In this paper, we introduce a hybrid, human centric interpretability framework that integrates rule-based modelling with SHAP-based feature attributions within a visual analytics framework and show the benefits for interpretability and interactivity through such techniques. We validate the framework on a case-study of Fishing vessel trajectories and demonstrate how this integrated approach reveals patterns and discrepancies that would not have been seen using a single approach alone.

Speaker: Bahavathy Kathirgamanathan (Fraunhofer IAIS)

23 Stretch-Compose: Semantic-Geometric Reasoning based Open Vocabulary Search and Retrieval of Objects in Dynamic Environments

Service robots operating in dynamic human environments must reliably locate objects that are moved, concealed, or completely novel. Current frameworks often assume static environments, failing to address these real-world uncertainties.
We present an open-vocabulary framework that combines spatial, semantic, and geometric reasoning to overcome these challenges. By unifying spatial cues about proximity and topology, semantic priors on typical placements, and geometric constraints that rule out infeasible locations, especially within concealed spaces, our approach finds objects even when they are relocated, hidden in drawers or cabinets, or first encountered through open-vocabulary queries. It also performs in-situ viewpoint planning to model relocated or unseen objects for manipulation and global scene-graph updates.
We validate our framework through extensive real-world trials on the Stretch SE3 mobile manipulator, evaluating search and retrieval in various conditions. Results demonstrate robust navigation (100%) and open-space detection (100%), with semantic-geometric reasoning reducing concealed space search time by 68% versus semantic-only approaches. The system achieves 80% detection in drawers, successfully capturing relocated objects through multi-view integration. Reasoning failed only for novel objects, requiring minimal user hints for correction. Although manipulation is constrained by the Stretch hardware, the cognitive stack consistently locates targets, demonstrating robust reasoning.
This work advances embodied AI by demonstrating how open-vocabulary, multi-modal reasoning enables robust object retrieval in dynamic and occluded environments. Implemented on a low-cost, compact mobile manipulator, our solution combines sophisticated cognitive capabilities with practical deployability, representing a significant step toward accessible service robots for everyday homes.

Speaker: Rohit Menon

menon25lamarr.pdf

24 Synthetic Data for Multi-Camera Multi-Object Tracking in Logistics

This abstract outlines my current research for my PhD thesis, focusing specifically on creating a synthetic dataset for multi-camera multi-object tracking (MCMOT) within logistics applications.

Motivation: Tracking moving assets such as trucks, trailers, or containers in logistics yards is crucial for developing digital twins, measuring key performance indicators, and enhancing operational efficiency. Effective MCMOT is essential yet challenging due to the presence of visually similar objects, frequent long-term occlusions, and extensive periods of low activity or inactivity.

Research Gap: Currently, there is a significant lack of open, logistics-specific datasets tailored for MCMOT applications. Most available datasets predominantly address urban or surveillance scenarios and rarely cover large-scale environments with overlapping fields of view for warehouses or cross-docking facilities. This absence severely restricts benchmarking, evaluation, and development of robust tracking algorithms for logistics scenarios, that don't rely heavily on re-identification.

Method: To address this gap, my research will involve the systematic creation of a large-scale synthetic dataset specifically designed for logistics yard scenarios. Characteristics of real-world warehouse and cross-dock environments will be realistically designed and simulated using Blender, building upon prior research on automated generation of labeled MCMOT datasets (https://ieeexplore.ieee.org/document/10710720). This dataset will focus explicitly on the tracking challenge, providing precise, automatically generated annotations to facilitate targeted experimentation and algorithmic development. Detection and segmentation challenges, e.g. addressing the sim-2-real gap, will be considered secondary to ensure primary focus remains on multi-camera tracking performance.

This targeted dataset creation effort aims to advance open research capabilities in logistics-specific MCMOT, enabling future development of more robust, efficient, and resource-aware tracking solutions.

Speaker: Christian Pionzewski (Fraunhofer IML)

LamarrPosterCP.pdf

25 Towards a Dataset of Realistic 3D Intralogistics Scenes for AI Applications

The advancement of artificial intelligence (AI) in intralogistics critically depends on the availability of realistic and diverse datasets. However, existing datasets in this domain often focus on narrow tasks such as object detection or activity recognition, lacking comprehensive three-dimensional (3D) representations of entire intralogistics systems. This paper addresses this gap by proposing a methodology for generating and formally describing synthetic 3D scenes of intralogistics environments, particularly warehouses, using a procedural content generation (PCG) approach. The proposed method constructs warehouse layouts by randomly placing and populating functional areas such as goods in, goods out, storage, packaging, and charging zones with realistic assets selected from domain specific libraries. Each generated scene is accompanied by a structured matrix describing asset positions, orientations, and types, facilitating downstream AI applications without needing to parse complex 3D model files. The approach was implemented using Nvidia’s Isaac Sim, producing a dataset of 100 diverse warehouse scenes comprising over 19,000 assets. The dataset's variability is confirmed through statistical metrics of capacity, workstations, and charging infrastructure. This foundational dataset aims to support a wide range of AI applications, from robot navigation and vision-based scene understanding to digital twin modeling and system simulation. Future work will include finding metrics to measure realism, finding learning-based approaches for the generation of the dataset, as well as incorporating further logistics processes and assets.

Speaker: Anike Murrenhoff

towards3Ddata.pdf

26 Towards Causal Understanding in Oncology: Structural Causal Modeling with Missing Data

Understanding causal relationships in oncology is critical for optimizing treatment strategies and generating testable biomedical hypotheses. We present CaDSIm (Causal Discovery with Simultaneous Imputation), a novel method for learning causal structures and associated Structural Equation Models (SEMs) from real-world data.

Our approach addresses three key objectives: Validation, Identification, and Counterfactual Reasoning. First, we validate CaDSIm by assessing its ability to recover known causal relationships in oncology. Second, we use it to uncover novel, potentially actionable dependencies among patient characteristics, tumor profiles, and treatment variables. Finally, we propose to leverage the learned causal model to answer counterfactual "what-if" questions, providing insights into treatment efficacy and patient outcomes.

Speakers: Michael Kamp, Osman Mian (The Lamarr Institute for Machine Learning and Artificial Intelligence)

Lamarr_Poster_final.pdf

27 Towards Universal Knowledge Graph embeddings

Cross-knowledge-graph (KG) learning is hindered because embeddings trained independently occupy incompatible vector spaces, while pre-merging KGs to enforce consistency is computationally infeasible at web scale. We present WHALE-embeddings, a continuously updated resource derived from Web Data Commons (~98B RDF triples across ~22M domains). By partitioning the corpus by website and training subgraphs independently with DECAL on HPC infrastructure, we obtain vectors for ~20.9B IRIs. To make these independently trained spaces interoperable, we introduce NAAS (Neural Adaptive Alignment Space) a model-agnostic, post-hoc alignment framework with two modes: NAASEA for entity alignment and NAASFT for iterative fine-tuning with KvsAll scoring. NAAS aligns local subgraphs and external KGs into a unified space without requiring graph fusion or retraining from scratch. Experiments on entity alignment and link prediction show that NAAS preserves strong downstream performance while enabling cross-KG nearest-neighbor search, disambiguation, and class expression learning. Together, WHALE-embeddings and NAAS provide a scalable path toward web scale, cross-domain representation learning and make the largest public KGE resource immediately usable across graphs.

Speakers: Duygu Ekinci (Paderborn University), Shivam Sharma

28 VISOR: VIsual Seizure Onset detection peRsonalized for epilepsy patients

The onset detection of epileptic seizures from multivariate Electroencephalogram (EEG) data is a challenging task. The variation in seizure patterns across patients and epilepsy types makes it particularly difficult to create a generic solution. Existing approaches indicate low recall due to their inability to capture complex seizure onset patterns. In this paper, we propose VISOR – a novel approach to detect the onset of epileptic seizures based on novel patient profiles and visual, personalized feature representations. VISOR leverages a vision transformer model to learn the spatio-temporal relationships between features, capture individual seizure propagation patterns, and perform seizure onset detection in a heterogeneous multi-patient dataset. Evaluation on a real-world dataset demonstrates that VISOR outperforms the state-of-the-art baselines by at least 5 percentage points for seizure onset detection in terms of the F1 score and indicates higher effectiveness for more complex patterns of propagating seizures.

Speaker: Uttam Kumar

29 Zero-shot Imputation with Foundation Inference Models for Dynamical Systems

Dynamical systems governed by ordinary differential equations (ODEs) serve as models for a vast number of natural and social phenomena. In this work, we offer a fresh perspective on the classical problem of imputing missing time series data, whose underlying dynamics are assumed to be determined by ODEs. Specifically, we revisit ideas from amortized inference and neural operators, and propose a novel supervised learning framework for zero-shot time series imputation, through parametric functions satisfying some (hidden) ODEs. Our proposal consists of two components. First, a broad probability distribution over the space of ODE solutions, observation times and noise mechanisms, with which we generate a large, synthetic dataset of (hidden) ODE solutions, along with their noisy and sparse observations. Second, a neural recognition model that is trained offline, to map the generated time series onto the spaces of initial conditions and time derivatives of the (hidden) ODE solutions, which we then integrate to impute the missing data. We empirically demonstrate that one and the same (pretrained) recognition model can perform zero-shot imputation across 63 distinct time series with missing values, each sampled from widely different dynamical systems. Likewise, we demonstrate that it can perform zero-shot imputation of missing high-dimensional data in 10 vastly different settings, spanning human motion, air quality, traffic and electricity studies, as well as Navier-Stokes simulations — without requiring any fine-tuning. What is more, our proposal often outperforms state-of-the-art methods, which are trained on the target datasets.Our pretrained model, repository and tutorials are available online.

Speaker: Patrick Seifner (University of Bonn)

FIM-Imputation_Patrick_Seifner_Lamarr_Scientific_Forum_Poster.pdf

30 AI-based sensor layout for predicting thermal deformations of CFRP machine tools

Using data-based approaches, accurate predictions of thermal deformations, which can significantly affect the quality of manufactured components, can be enabled. However, a sufficient amount of data with maximised information content is necessary for efficient training. In this paper, an approach for optimising sensor configurations for predicting thermal deformations is presented. From initially 300 temperature sensors, the number of required sensors was significantly reduced while maintaining predictive accuracy. Furthermore, a pattern for sensor placement was identified, providing the potential for an efficient sensor layout that enables cost-effective data acquisition and improved monitoring of machining and wear progression of machine tool components.

Speaker: Felix Finkeldey

31 Database-driven Automation of Machine Learning Reconstruction for Imaging Air Cherenkov Telescopes

For more than two decades, the MAGIC telescopes continuously accumulate
significant amounts of data. However, the analysis of this data poses critical
problems due to its volume exceeding existing data curation capacities. This
criticality induces the demands for the utilization of AI methods to enhance and
accelerate the analysis process. Thus, MAGIC utilizes random forests for an ac-
celerated and robust reconstruction of the energy and direction of the measured
particles.
Consequently, efficient analysis performed with the respective AI methods re-
quires the development of a tool that ensures traceability as well as repro-
ducibility. Therefore, we present the database-driven tool autoMAGIC, capa-
ble of coordinating the use of random forests for large-scale datasets. Based
on the analysis specifications, autoMAGIC runs the respective tools for choos-
ing suitable training data, training and testing the random forest, and storing
the outputs for further processing over multi-year datasets. Furthermore, we
present long-term lightcurves performed with autoMAGIC, demonstrating the
use of autoMAGIC to acquire labor-intensive AI-based results efficiently.

Speakers: Mr Felix Wersig (TU Dortmund University / LAMARR Institut), Mr Luca Di Bella (TU Dortmund University / LAMARR Institut)

LAMARR_Scientific_Forum_Abstract-3.pdf

32 Reduction of data labeling effort by stability evaluation using drift detection techniques for milling operations

Designing stable milling operations is crucial to ensure a high surface quality of the machined workpieces and reduce rejects during production. Stability lobe diagrams can be used to identify stable conditions. Analytical approaches or simulation techniques can be used to reduce the experimental effort for stability evaluation for different process parameter values. However, complex cause-effect relationships, such as concept drift due to tool wear, and simplifying assumptions required to ensure sufficient simulation run-times limit their applicability. In contrast, machine learning models can provide real-time predictions of process characteristics based on different process parameter configurations with high accuracy and generalization. However, a large number of experiments are still necessary to establish the required database. In this context, labeling the data, i.e., evaluating the resulting stability for each parameter configuration, can be time-consuming. To this end, sensor data collected during the process can be used to automate the stability assessment. However, sensors that can be efficiently integrated into the working area without interfering with the experimental setup, e.g., acoustic emission sensors, are often susceptible to noise, which makes algorithmic analysis challenging. In this paper, a framework for automatic evaluation of milling stability based on statistical tests using time series data acquired by acoustic emission sensors is presented. The proposed framework also considers model uncertainty and has been validated on controlled synthetic and noisy real data sets. In addition, insightful analyses of the model hyperparameters are given for efficient model performance.

Speaker: Shubham Gupta

3:30 PM Grab your coffee and head to the next session

Transfer from first to ground floor

The Lamarr Moonshot: Spearheading European, open-source Foundation Models

Evolving and growing together: The Lamarr Institute is shaping the future of AI. To this end, the Lamarr Institute is advancing the research and development of Lamarr-driven data-centric, multi-lingual foundation models with vertical projection into the Lamarr Research Areas.

Join in to learn more about the Lamarr Moonshot and how you can contribute to move the Lamarr vision forward.

Convener: Mehdi Ali

33 Closing of day one

The Lamarr Scientific Forum is rounding off the first day with a closing and all information needed on dinner plans.

5:30 PM [Time for Transfer and Check-In]

7:00 PM Dinner

L'Osteria Bonn Webpage
Google Maps

Thu, September 4

8:30 AM Arrival (optional)

Time to arrive at the venue. There will be coffee and some small snacks.

34 Welcome

We begin program day number 2 with a short look back at the previous day and ahead at today's program.

Lamarr Area Updates: Session 2

Working and growing together: On each day, 5 of the 10 Lamarr areas will provide the attendees with an update on its respective strategic focus, breakthrough results and opportunities for cooperation.

35 Area Presentation: Human-Centered Systems

Our research in Human-Centered AI focuses on enabling domain experts to actively guide and interpret machine learning (ML) processes through interactive, knowledge-driven methods. We develop visual analytics (VA) techniques that support expert involvement in both the construction and interpretation of ML models, with the goal of improving transparency, trust, and alignment with human reasoning.

A key strand of our work involves supporting domain experts in transforming raw temporal and spatial data into semantically meaningful, episode-based representations. These enriched representations enable the training of ML models whose behavior can be more easily understood and validated. In parallel, we have explored methods for generating domain-aware synthetic instance neighborhoods to support localized model explanations, ensuring the plausibility and realism of instance-based reasoning.

The core of our current work centers on RuleSense, a visual analytics system designed to support expert-driven exploration of model logic through rule-based representations. RuleSense enables users to analyze and make sense of large rule sets extracted from trained models—such as random forest classifiers and regressors—by providing aggregated views of how feature values relate to predicted outcomes. The system allows experts to investigate whether a model’s internal logic is consistent with domain knowledge and to identify unexpected or potentially flawed decision patterns.

We demonstrate RuleSense across several use cases: from COVID-19 incidence prediction to movement pattern recognition in maritime data. In the life sciences domain, we apply it to a regression model predicting the medical potency of chemical compounds. By encoding extracted rules and applying topic modeling techniques, we help experts discover potentially interesting feature combinations associated with high predicted outcomes—an approach that supports hypothesis generation in early-stage drug discovery.

Together, these efforts contribute to the broader goal of designing AI systems that are not only technically effective but also interpretable, controllable, and meaningful to the human decision-makers who rely on them.

Speaker: Natalia Andrienko (Fraunhofer Institute IAIS)

36 Area Presentation: AI in Life Sciences and Health

The Lamarr interdisciplinary research area "AI in Life Sciences and Health" will provide an overview of its organization and scientific focal points in the life sciences including drug discovery, medicine, and health. We will introduce research groups participating in this area and key collaborations with external partners and institutions. In addition, recent research progress will be discussed.

The introductory presentation will be followed by highlighting a recent project in the XAI area, generating the computational explanations for equivariant diffusion models used for molecular design. Predictions of diffusion models are hard to rationalize, charting new territory in XAI, beyond the analysis of model-internal weights and gradients. This work reports the first explanations for diffusion models at the molecular level of detail.

Speakers: Jürgen Bajorath, Elena Xerxa (University of Bonn), Andrea Mastropietro (University of Bonn)

37 Area Presentation: Embodied AI

This talk provides an overview of recent research in the area of Embodied AI at Lamarr. Embodied Artificial Intelligence refers to AI that is embedded in physical systems, such as robots, and can interact with the surroundings. In contrast to classic Machine Learning in robotics, embodied AI encapsulates all aspects of interacting and learning in an environment: from perception, via understanding, reasoning, and planning to execution respectively manipulation. This field brings together various disciplines, including computer vision, environment modeling, prediction, planning, and control, reinforcement learning, physics-based simulation, and robotics.

Throughout the presentation, we will explore the goals and recent achievements in Embodied AI, showcase some breakthrough results, and discuss opportunities to collaborate in this vibrant area of research. Join us to learn more about how Embodied AI is shaping the future of intelligent systems and how you can get involved.

Speakers: Sven Behnke (Universität Bonn), Julian Eßer

38 Area Presentation: NLP

Over the past year, the Natural Language Processing (NLP) research area at the Lamarr Institute has made significant strides toward building more robust, context-aware, and aligned language technologies. This talk will provide an overview of key developments in this area and our future plans. We will highlight flagship publications, newly funded projects, strategic collaborations and activities to Lamarr’s NLP research at the forefront of European AI innovation. Looking ahead, the presentation will outline our roadmap for future research, including scalable reasoning and evaluation methods, value-grounded language generation, and integrating feedback into model development. Anchored in the “Triangular AI” paradigm, our approach continues to balance algorithmic excellence, high-quality data curation, and societal impact. This session aims to foster dialogue on shaping responsible NLP research and invites continued collaboration across disciplines and sectors.

In addition, two selected research projects will be presented, sharing insights into multi-hop reasoning for question answering with hyperbolic representations and superalignment with dynamic human values.

Speakers: Lucie Flek (University of Bonn), Akbar Karimi (University of Bonn), Florian Mai (University of Bonn)

39 Area Presentation: Industry and Production

The Industry and Production research area focuses on the integration of artificial intelligence and machine learning (ML) into production technology. The main objectives are to ensure consistent product quality while minimizing the use of resources such as machine time, tools, materials and energy. This presentation provides an overview of the main research topics of the area, which are aligned with the overarching strategy of triangular AI, bringing together data-driven observations, physically-based models and expert knowledge. Through these scientific efforts, the research area aims to contribute to a future in which production characteristics and system states are analyzed automatically. This will enable real-time optimization and continuous improvement of ML models as new data becomes available during operation.

Spotlight: "Cluster-based Prediction of Chatter Vibrations in Milling Operations"

Speaker: Felix Finkeldey

11:45 AM Lunch

Lunch for all Lamarr members is served in the foyer of the ground floor. We encourage you to use the opportunity to network with fellow Lamarr staff.

Poster Session: Session 2

Learning and growing together: The poster session offers Lamarr members with a platform to showcase their research projects. Lamarr members across all research areas, working groups and levels of hierarchy and experience are asked to join the open exchange and discuss the presented work.

How to protect your presentation Materials.pdf

40 AALF: Almost Always Linear Forecasting

Recent works for time-series forecasting more and more leverage the high predictive power of Deep Learning models.
With this increase in model complexity, however, comes a lack in understanding of the underlying model decision process, which is problematic for high-stakes application scenarios. At the same time, simple, interpretable forecasting methods such as ARIMA still perform very well, sometimes on-par, with Deep Learning approaches. We argue that simple models are good enough most of the time, and that forecasting performance could be improved by choosing a Deep Learning method only for few, important predictions, increasing the overall interpretability of the forecasting process.
In this context, we propose a novel online model selection framework which learns to
identify these predictions. An extensive empirical study on various real-world datasets shows that our selection methodology performs comparable to state-of-the-art online model selections methods in most cases while being significantly more interpretable. We find that almost always choosing a simple autoregressive
linear model for forecasting results in competitive performance, suggesting that the need for opaque black-box models in time-series forecasting might be smaller than recent works would suggest.

Speaker: Matthias Jakobs (TU Dortmund)

aalf_poster.pdf

41 Adaption of process monitoring motivated by a combination of rejection ensembles

In machining processes, measurement data collected during operation can provide valuable insights into the final quality of the manufactured components. This enables both the reduction of unnecessarily long process chains and the real-time adaptation of process parameters. Machine learning models trained on this data can capture and predict complex process characteristics. However, acquiring comprehensive measurement data is often expensive and technically challenging, making resource-efficient deployment strategies essential. One promising approach is the use of resource-aware rejection ensembles. By categorizing measurement data into low-cost, contactless measurements; medium-cost, integrated measurements; and complex but highly reliable measurements, it becomes possible to optimize rejection strategies, for example, through the targeted selection of measurement devices based on the uncertainty profiles of different predictive models.

Speaker: Jim Bergmann

42 AI and Astroparticle Physics with Neutrinos

Over the last decade, AI-algorithms have become a standard tool for data analysis in astroparticle physics. While these efforts were pioneered by the use of ensemble methods for event selection in IceCube, the capabilities of AI in the context of neutrino astronomy have been exemplified by the detection of neutrinos from the Milky Way (also by the IceCube collaboration). This detection was enabled by the use of deep neural networks, which allow for a more precise reconstruction of the neutrino trajectory and a more sophisticated event selection. At the same time efforts are underway to also leverage the capabilities of GNNs and transformer models for the event reconstruction in neutrino astronomy. This poster will provide an overview over the state-of-the-art with respect to the utilization of AI in astroparticle physics, with a dedicated focus on neutrino astronomy.

Speaker: Tim Ruhe (TU Dortmund)

43 Anytime YOLO - Early exits for interruptable object detection

Anytime algorithms can be interrupted before completion while still delivering an intermediate result. This is a desirable property for embedded systems where timing is critical, such as object detection in cyber-physical-systems. However it is generally neither supported by models nor inference frameworks. To enable a model to be anytime, early-exits can be added to the network, which allow an earlier branching off from intermediate layers. We present initial explorations on how to formally define anytime quality, modify the well-known YOLO object detection model to be anytime, in regards to how early-exits can be added and how the architecture can be modified, and the current feasibility of anytime inference.

Speaker: Daniel Kuhse

anytime yolo poster.pdf

44 Approximately Pareto-optimal Solutions for Bi-Objective k-Clustering Problems

As a major unsupervised learning method, clustering has received a lot of attention over multiple decades. The various clustering problems that have been studied intensively include, e.g., the k-means problem and the k-center problem. How- ever, in applications, it is common that good clusterings should optimize multiple objectives (e.g., visualizing data on a map by clustering districts into areas that are both geographically compact but also homogeneous with respect to the data). We study combinations of different objectives, for example optimizing k-center and k-means simultaneously or optimizing k-center with respect to two different metrics. Usually these objectives are conflicting and cannot be optimized simul- taneously, making it necessary to find trade-offs. We develop novel algorithms for approximating the set of Pareto-optimal solutions for various combinations of two objectives. Our algorithms achieve provable approximation guarantees and we demonstrate in several experiments that the approximate Pareto front contains good clusterings that cannot be found by considering one of the objectives separately.

Speaker: Sarah Sturm (University of Bonn)

45 Chi: Symmetry Understanding of 3D Shapes via Chirality Disentanglement

Chirality information (i.e., information that allows distinguishing left from right) is ubiquitous for various data modes in computer vision, including images, videos, point clouds, and meshes. Contrary to symmetry, for which there has been a lot of research in the image domain, chirality information in shape analysis (point clouds and meshes) has remained underdeveloped. Although many shape vertex descriptors have shown appealing properties (e.g. robust to rigid-body pose transformations), they are not able to disambiguate between left and right symmetric parts. Considering the ubiquity of chirality information in different shape analysis problems and the lack of chirality-aware features within current shape descriptors, developing a chirality feature extractor becomes necessary and urgent. In this paper, based on the recent framework Diff3f, we proposed an unsupervised chirality feature extraction pipeline to decorate shape vertices with chirality-aware information, extracted from 2D foundation models. Quantitative and qualitative results of various experiments and downstream tasks include left-right disentanglement, shape matching, and part segmentation conducted on a variety of datasets proving the effectiveness and usefulness of our extracted chirality features.

Speakers: Weikang Wang (Learning and Optimisation for Visual Computing Group, University of Bonn), Tobias Weißberg (Learning and Optimisation for Visual Computing Group, University of Bonn)

chi_lamarr_4_1.pdf

46 CINEMETRIC: A Framework for Multi-Perspective Evaluation of Conversational Agents using Human-AI Collaboration

Despite advances in conversational systems, the evaluation of such systems remains a challenging problem. Current evaluation paradigms often rely on costly homogeneous human annotators or oversimplified automated metrics, leading to a critical gap in socially aligned conversational agents, where pluralistic values (i.e., acknowledging diverse human experiences) are essential to reflect the inherently subjective and contextual nature of dialogue quality. In this paper, we propose CINEMETRIC, a novel framework that operationalizes pluralistic alignment by leveraging the perspectivist capacities of large language models. Our approach introduces a mechanism where LLMs simulate a diverse set of evaluators, each with distinct personas constructed by matching real human annotators to movie characters based on both demographic profiles and annotation behaviors. These role-played characters independently assess subjective tasks, offering a scalable and human-aligned alternative to traditional evaluation. Empirical results show that our approach consistently outperforms baseline methods, including LLM as a Personalized Judge, across multiple LLMs, showing high and consistent agreement with human ground truth. CINEMETRIC improves accuracy by up to 20% and reduces mean absolute error in toxicity prediction, demonstrating its effectiveness in capturing human-like perspectives. We further extend CINEMETRIC with a causal analysis pipeline to identify how latent factors such as cultural background and personality traits cause systematic differences in toxicity perception across perspectives, bridging pluralistic alignment with interpretability.

Speaker: Mr Vahid Sadiri Javadi (University of Bonn)

LAMARR - Poster - CINEMETRIC.pdf

47 Constructive Empiricism for Explainable AI

We explore what it means to build a scientific "theory" of a black-box model, drawing on van Fraassen's Constructive Empiricism (CE), and demonstrate how such a theory can be used for explainable AI (XAI).
A scientific theory is more than just an explanation: it not only has value in its own right, but also serves as a robust framework for answering different questions.
According to CE, a theory must be both empirically adequate (i.e., accurate with respect to observed data) and shaped by pragmatic virtues, such as user preferences. These criteria align closely with the needs of XAI, which require fidelity and comprehensibility.
We turn CE's core notion of empirical adequacy into three concrete criteria: consistency, sufficient predictive performance, and algorithmic adaptability. As a proof of concept we develop the Constructive Box Theorizer (CoBoT) algorithm within this framework.

Speaker: Sebastian Müller (University of Bonn)

semueller_CE4XAI_poster.pdf

48 Deep Learning Applications in Radio Interferometry

In recent years, machine learning and deep learning have revolutionized data analysis across various fields, including particle physics and medical imaging. However, their potential in radio interferometry—a technique used to study the universe through arrays of radio telescopes—remains underexplored. The radionets-project has been pioneering the use of deep learning methods to process data from radio interferometers for over five years. These techniques promise significant improvements in performance and sensitivity, particularly for next-generation instruments like MeerKat and SKA.

Our recent publications demonstrates how deep learning can enhance imaging workflows by reconstructing high-quality images from complex datasets generated by interferometers such as VLA, VLBA, and ALMA. While initial results on simulated data have shown great promise, ongoing efforts are focused on adapting these models to real observational data from MeerKat. This poster will provide an overview of our reconstruction approach, describe the simulation pipeline we developed to train deep learning models effectively, and highlight future opportunities for applying these techniques in analyzing radio telescope data.

Speaker: Dr Kevin Schmitz

radionets_lamarr_bonn.pdf

49 Efficient Manipulation-Enhanced Semantic Mapping With Uncertainty-Informed Action Selection

Service robots operating in cluttered human environments such as homes, offices, and schools cannot rely on predefined object arrangements and must continuously update their semantic and spatial estimates while dealing with possible frequent rearrangement. Identifying all objects in cluttered, occlusion-heavy environments, such as shelves, requires selecting informative viewpoints and performing targeted manipulations to reduce uncertainty regarding object locations, shapes, and categories. We present a unified and manipulation-enhanced, semantic mapping framework that addresses this challenge as a partially observable Markov decision process (POMDP), whose high-dimensional belief is represented by an evidential, metric-semantic grid map. To efficiently reason about occlusions, and manipulation effects, we propose Calibrated Neural-Accelerated Belief Updates (CNABUs): a neural network–based belief propagation model that produces confidence-calibrated predictions for unknown areas. Uncertainty estimates from Dirichlet distributions (for semantic predictions) and Beta distributions (for occupancy) guide active sensor placement via reinforcement learning–based next-best view planning and object manipulation via an uncertainty-informed push strategy targeting occlusion-critical objects. By focusing on areas of limited knowledge and selecting actions with high expected information gain, our method minimizes unwanted object displacement and dropping. Our planner substantially improves map completeness and accuracy compared to existing approaches while reducing planning time by 95%. Our approach successfully transfers to real-world cluttered shelves in a zero-shot fashion, demonstrating its robust real-world applicability. This work has been accepted for the Robotics Science and Systems Conference 2025 and an extension is currently under submission for HUMANOIDS 2025.

Speaker: Nils Dengler (University of Bonn)

Dengler_Lamarr_Poster.pdf

50 First Investigation of Deep Learning for Intraoperative Gauze Segmentation in Minimally Invasive Abdominal Surgery

Surgical gauze is an essential part of surgical procedures, which is primarily used for controlling bleeding and absorbing bodily fluids. The post-surgical retention of gauze can lead to serious complications in the patient’s health and necessitate additional surgery for gauze removal. In the wake of data scarcity, the research on gauze segmentation on the real-world surgical data remains underexplored. In this work, we investigate the use of deep learning methods for gauze segmentation in robot-assisted minimally invasive abdominal surgeries, utilizing an in-house surgical dataset prepared at a university hospital. The training data reflects a realistic surgical setting and extensive diversity in spatial, morphological, and visual attributes of three different gauze categories. We have investigated prevalently used segmentation architectures, including CNN-based, transformer-based, and hybrid architectures, to provide a proof-of-concept for gauze segmentation in a realistic setting. Besides, we investigate the influence of additional sub-optimally annotated, auto-tracked segmentation masks to address the bottleneck of data scarcity and performance optimization. Our results demonstrate the efficacy of real-world data to counter the main challenge reported by prior works — the trade-off between blood presence and gauze detection. The incorporation of auto-track annotations enables performance enhancements, particularly in generic cases. The integration of effective segmentation approaches will benefit robot-guided surgical procedures and various downstream applications by providing a precise delineation of foreign objects, enhancing patient safety and surgical outcomes.

Speaker: Priya Priya (Fraunhofer IAIS, University of Bonn)

DSAA_Poster_Lamarr.pdf

51 From Prediction to Insight: Visual Analytics for Understanding Compound Potency Models

Providing clear explanations is crucial in interdisciplinary research fields like bioinformatics where non-experts in machine learning (ML) must understand model decisions to foster trust in the system. Interactive visualisation can help in enabling the active exploration of model behaviour. In this paper,we present an approach to interpreting compound potency predictions by using RuleSense, a visual analytics system that integrates rule-based modeling, topic modeling, and interactive visual tools to support deeper interpretation of predictive models. RuleSense groups related rules into coherent topics, identifies key feature patterns, and allows users to trace how these patterns contribute to predictions. This helps to reduce the complexity of the model and helps to gain understanding on the most meaningful parts of a model. We demonstrate the approach in the context of compound potency prediction—a critical task in drug discovery—showing how RuleSense reveals relationships between chemical structure and activity, supports hypothesis generation, and bridges the gap from prediction to scientific insight.

Speaker: Dr Tiago Janela

From_prediction_to_insight_poster.pdf

52 Hierarchical Vector Quantization for Unsupervised Action Segmentation

In this work, we address unsupervised temporal action segmentation, which segments a set of long, untrimmed videos into semantically meaningful segments that are consistent across videos. While recent approaches combine representation learning and clustering in a single step for this task, they do not cope with large variations within temporal segments of the same class. To address this limitation, we propose a novel method, termed Hierarchical Vector Quantization (HVQ), that consists of two subsequent vector quantization modules. This results in a hierarchical clustering where the additional subclusters cover the variations within a cluster. We demonstrate that our approach captures the distribution of segment lengths much better than the state of the art. To this end, we introduce a new metric based on the Jensen-Shannon Distance (JSD) for unsupervised temporal action segmentation. We evaluate our approach on three public datasets, namely Breakfast, YouTube Instructional and IKEA ASM. Our approach outperforms the state of the art in terms of F1 score, recall and JSD.

Speaker: Federico Spurio (University of Bonn)

Spurio_HVQ.jpg

Spurio_Lamarr-Scientific-Poster_A0.pdf

53 In-Training Defenses against Emergent Misalignment in Language Models

Fine-tuning lets practitioners repurpose aligned large language models (LLMs) for new domains, yet recent work reveals emergent misalignment (EMA): Even a small, domain-specific fine-tune can induce harmful behaviors far outside the target domain. Even in the case where model weights are hidden behind a fine-tuning API, this gives attackers inadvertent access to a broadly misaligned model in a way that can be hard to detect from the fine-tuning data alone. We present the first systematic study of in-training safeguards against EMA that are practical for providers who expose fine-tuning via an API. We investigate four training regularization interventions: (i) KL-divergence regularization toward a safe reference model, (ii) $\ell^2$ distance in feature space, (iii) projecting onto a safe subspace (SafeLoRA), and (iv) interleaving of a small amount of safe training examples from a general instruct-tuning dataset. We first evaluate the methods’ emergent misalignment effect across four malicious, EMA-inducing tasks. Second, we assess the methods’ impacts on benign tasks. We conclude with a discussion of open questions in emergent misalignment research.

Speaker: David Kaczér (University of Bonn)

Copy of poster-template-lamarr-1.pdf

54 Jailbreaking LLMs Without Gradients or Priors: Effective and Transferable Attacks

Large Language Models (LLMs) remain vulnerable to adversarial jailbreaks, yet existing attacks rely on handcrafted priors or require white-box access for gradient propagation. We show that token-level iterative optimization can succeed without gradients and introduce RAILS (RAndom Iterative Local Search), a simple yet effective method using only model logits with a query budget comparable to gradient-based approaches. To improve attack success rates (ASRs), we incorporate a novel auto-regressive loss and history buffer-based candidate selection for few-shot attacks, achieving near 100\% ASRs on robust open-source models. By eliminating token-level gradients, RAILS enables cross-tokenizer attacks. Notably, attacking ensembles of diverse models significantly enhances adversarial transferability, as demonstrated on closed-source systems such as GPT-3.5, GPT-4, and Gemini Pro. These findings demonstrate that handcrafted priors and gradient access are not necessary for successful adversarial jailbreaks, highlighting fundamental vulnerabilities in current LLM alignment.

Speaker: Zhakshylyk Nurlanov (Learning and Optimisation for Visual Computing Group, University of Bonn)

lamarr_llm_attack_poster.pdf

55 JSON is all you need (for visual symbolic planning)

Traditional Task test and Motion Planning (TAMP) systems integrate physics simulators for motion planning with discrete symbolic models for task planning. However, because these symbolic models are not derived from data, they must be meticulously handcrafted, requiring manually designed classifiers to bridge the gap with the physics simulator. This process is both resource-intensive and constrained to the specific domain for which it was engineered, limiting scalability and adaptability. Due to their extensive training on heterogeneous data, Visual Language Models (VLMs) are well suited for TAMP-like problems in the open-world. However, they have limited real-world grounding and planning capabilities. Therefore, recent efforts have been made to integrate VLMs with classical planning for long-horizon reasoning. However, they still depend on task-specific solutions, e.g. describing all possible objects in advance, and symbolic action models. We propose a novel framework that does not require either. It leverages VLMs to retrieve symbolic representations directly from images, based on only lifted predicates. To integrate with classical planning, we extend the heuristic-free Width-Based search algorithm to handle probabilistic representations such as one generated by VLMs. We evaluate our system using the PddlGym environment and the Problem Description Generation Dataset.

Speaker: Sami Azirar

use_case_diagram_ssr_parsing.png

56 Little Is Enough: A Privacy-Preserving Framework for Healthcare Models

In the healthcare domain, sensitive patient data is inherently decentralized across institutions and cannot be centralized due to strict privacy regulations. Federated learning offers a collaborative model training without explicitly sharing patient data by communicating model parameters or soft labels. These approaches, however, are still vulnerable to privacy leakage and often limit model flexibility to those trainable by gradient-based methods. We present a novel federated co-training (FEDCT) method that enhances privacy substantially by exchanging only hard (definitive) labels on a shared public unlabeled dataset. Participating healthcare institutions collaboratively generate consensus labels, which are used as pseudo-labels to train local models. This approach not only empirically improves resistance to membership inference attacks but also supports a wider range of models, including interpretable and non-differentiable algorithms like decision trees and ensemble methods. FEDCT is particularly suited for healthcare use cases such as distributed radiology analysis or clinical data modeling over several hospitals, where model interpretability, privacy, and communication efficiency are paramount. Our theoretical analysis and empirical evaluations on both general and medical datasets demonstrate that FEDCT achieves high model performance with significantly improved privacy guarantees, facilitating secure and practical federated learning in sensitive healthcare environments.

Speaker: Amr Abourayya (Trustworthy AI)

57 LLM Value Alignment

Social sciences define values as preferred behaviors or outcomes that motivate an individual's actions or judgments.
While LLMs often reflect biases from their training data, it remains unclear what values underlie their generation processes, and whether such internal value systems can be measured or modified.
In this paper, we investigate whether fine-tuning can steer a model’s internal moral preferences and whether such changes manifest in downstream behavior.
Building on a taxonomy of 20 human values, we fine-tune models using two approaches: supervised fine-tuning (SFT) with scalar value ratings in a survey; and direct preference optimization (DPO) with contrastive sentence pairs.
Each method downgrades a target value while keeping others fixed.
We evaluate models on moral judgments in the Am I The Asshole subreddit, using GPT-labeled examples with high vs. low value standards.
We measure both prediction change rate and directional consistency with expected value shifts.
Results show that SFT is more effective than DPO at inducing value-aligned behavioral changes, especially for values with sufficient evaluation data. These findings suggest that value-specific instruction tuning offers a promising path for aligning LLMs' moral behavior.

Speaker: Shangrui Nie (Bonn-Aachen International Center for Information Technology (b-it))

58 MCBench: A Benchmark Suite for Monte Carlo Sampling Algorithms

In this paper, we present MCBench, a benchmark suite designed to assess the quality of Monte Carlo (MC) samples. The benchmark suite enables quantitative comparisons of samples by applying different metrics, including basic statistical metrics as well as more complex measures, in particular the sliced Wasserstein distance and the maximum mean discrepancy. We apply these metrics to point clouds of both independent and identically distributed (IID) samples and correlated samples generated by MC techniques, such as Markov Chain Monte Carlo or Nested Sampling. Through repeated comparisons, we evaluate test statistics of the metrics, allowing to evaluate the quality of the MC sampling algorithms.
Our benchmark suite offers a variety of target functions with different complexities and dimensionalities, providing a versatile platform for testing the capabilities of sampling algorithms. Implemented as a Julia package, MCBench enables users to easily select test cases and metrics from the provided collections, which can be extended as needed. Users can run external sampling algorithms of their choice on these test functions and input the resulting samples to obtain detailed metrics that quantify the quality of their samples compared to the IID samples generated by our package. This approach yields clear, quantitative measures of sampling quality and allows for informed decisions about the effectiveness of different sampling methods.
By offering such a standardized method for evaluating MC sampling quality, our benchmark suite provides researchers and practitioners from many scientific fields, such as the natural sciences, engineering, or the social sciences with a valuable tool for developing, validating and refining sampling algorithms.

Speaker: Zeyu Ding (TU Dortmund)

MCBench, Zeyu Ding.pdf

59 Pairwise-TD: A Bellman Operator for Relative Value Learning

Reinforcement learning traditionally learns absolute state values, estimating how good a particular situation is in isolation. Yet in both biological systems and practical decision-making, what often matters is not the absolute value of a state, but how it compares to alternatives. Motivated by empirical findings in neuroscience, we introduce \textbf{Pairwise-TD}, a novel framework that learns \emph{value differences} directly.
Our method defines a new pairwise Bellman operator that estimates the relative value $\Delta(s_i, s_j) = V(s_i) - V(s_j)$, bypassing the need to ever compute $V(s)$ explicitly. We prove that this operator is a $\gamma$-contraction in a structured function space, ensuring convergence to a unique fixed point. Pairwise-TD integrates naturally into on-policy actor-critic methods and enables exact recovery of Generalized Advantage Estimation (GAE) using only pairwise differences. Hereby, we derive a pseudo-value approach that yields an unbiased policy gradient estimator despite the absence of an explicit value baseline. To address pair-wise comparisons in episodic environments with terminal states, we introduce a principled scheme for computing Bellman targets using only observable quantities, ensuring correct learning even when episode lengths vary. Finally, we present a lightweight neural network architecture that enforces antisymmetry via a shared encoder and linear projection, further improving the structure of our relative value function. Together, these contributions offer a biologically inspired, practically effective, and theoretically grounded alternative to traditional value learning.

Speaker: Marc Höftmann

relative_value_learning_poster.pdf

60 Polyra Swarms

In this poster, I show Polyra Swarms, a novel approach to machine learning that shifts focus from function approximation to shape approximation. While these swarms are still less developed, I show that they can still hold their own when compared to neural networks and on some tasks outperform them. I also present an automated abstraction mechanism that enhances generalization and interpretability by simplifying the learned swarm structure. By operating on principles fundamentally different from neural networks, Polyra Swarms offer new perspectives and open up fresh research directions in the design of learning systems.

Speaker: Simon Klüttermann

polyra___poster.pdf

61 Predicting machining stability with a quantum regression model

In this article, we propose a novel quantum regression model by extending the Real-Part Quantum SVM. We apply our model to the problem of stability limit prediction in milling processes, a key component in high-precision manufacturing. To train our model, we use a custom data set acquired by an extensive series of milling experiments using different spindle speeds, enhanced with a custom feature map. We show that the resulting model predicts the stability limits observed in our physical setup accurately, demonstrating that quantum computing is capable of deploying ML models for real-world applications.

Speaker: Loong Kuan Lee (Fraunhofer IAIS)

62 State-Transition-Aware Anomaly Detection Under Concept Drifts

Detecting temporal abnormal patterns over streaming data is challenging due to volatile data properties and the lack of real-time labels. The abnormal patterns are usually hidden in the temporal context, which cannot be detected by evaluating single points. Furthermore, the normal state evolves over time due to concept drifts. A single model does not fit all data over time. Autoencoders are recently applied for unsupervised anomaly detection. Autoencoders have recently been applied for unsupervised anomaly detection. However, they are trained on a single normal state and usually become invalid after distributional drifts in the data stream. This paper uses an Autoencoder-based approach STAD for anomaly detection under concept drifts. In particular, we propose a state-transition-aware model to map different data distributions in each period of the data stream into states, thereby addressing the model adaptation problem in an interpretable way. In addition, we analyzed statistical tests to detect the drift by examining the sensitivity and powers. Furthermore, we present considerable ways to estimate the probability density function for comparing the distributional similarity for state transitions. Our experiments evaluate the proposed method on synthetic and real-world datasets. While delivering comparable anomaly detection performance as the state-of-the-art approaches, STAD works more efficiently and provides extra interpretability. The insightful analysis of optimal hyperparameters for efficient model training and adaptation is also discussed. Our ongoing research aims to develop the work further.

Speakers: Shubham Gupta, Mr Daniil Kaminskyi (TU Dortmund University)

63 Sustainable Model Selection via Meta-Learning

While many have analyzed the resource efficiency of trained models, an important question remains: How can one be sustainable and resource-aware during AI development, or in other words, when looking for a suitable model to train on a specific learning task? AutoML can help with finding well-performing models on given data, however these frameworks overly focus on predictive quality and neglect performance trade-offs. To answer the calls for sustainable and green AutoML, this poster presents compositional meta-learning (CML). As an explainable approach to model selection, it allows to not only acknowledge multiple performance aspects but also to consider user priorities. With strong generizability and demonstrated effectiveness for time series forecasting and tabular data classification, CML is an important extension to the meta-learning idea and facilitates resource-aware ML.

Speaker: Raphael Fischer (TU Dortmund University)

raml_poster.pdf

64 Synthetic Data from Physics Simulations: A Viable Alternative for Pallet Activity Recognition?

Pallets are one of the most important load carriers for international supply chains. Yet, continuously tracking activities such as driving, lifting or standing along their life cycle is hardly possible. As part of a preliminary project, it was shown that it is possible to develop a prediction model for pallet activities using data from inertial measurements units mounted on a pallet. A significant challenge in the development of the prediction model is the manual recording and annotation of processes, which significantly restricts the available data. The utilisation of synthetic data derived from physics simulations provides a potential solution to the challenges posed by the scarcity of data and the inability to identify a comprehensive range of processes. To validate this approach, data is recorded in real intralogistics environments and simulation models are created that mimic these real processes. The simulation models are then used to generate synthetic data. To record the real data, a new sensor board was developed, which was adapted for recording in industrial environments. The quality of the synthetic data is then evaluated by comparing it with the recorded data.

Speakers: Alexander Krooß (Fraunhofer IML), Jan Jäkel, Julian Brandt

Pallet to Simulation - Synthetic Data from Physics 2.pdf

65 The Anatomy of Evidence: An Investigation Into Explainable ICD Coding

Automatic medical coding has the potential to ease documentation and billing processes. For this task, transparency plays an important role for medical coders and regulatory bodies, which can be achieved using explainability methods. However, the evaluation of these approaches has been mostly limited to short text and binary settings due to a scarcity of annotated data. Recent efforts by Cheng et al. (2023) have introduced the MDACE dataset, which provides a valuable resource containing code evidence in clinical records. In this work, we conduct an in-depth analysis of the MDACE dataset and perform plausibility evaluation of current explainable medical coding systems from an applied perspective. With this, we contribute to a deeper understanding of automatic medical coding and evidence extraction. Our findings reveal that ground truth evidence aligns with code descriptions to a certain degree. An investigation into state-of-the-art approaches shows a high overlap with ground truth evidence. We propose match measures and highlight success and failure cases. Based on our findings, we provide recommendations for developing and evaluating explainable medical coding systems.

Speaker: Katharina Beckh

ACL2025-Find-poster-Anatomy.pdf

66 Towards Human Motion Generation for Industrial Simulations

Human motion is of interest to industrial simulations for process optimisation, ergonomic evaluation and visualisation of digital-twin environments. Furthermore, it is of interest for simulation-based reinforcement learning in human-robot interaction and humanoid robotics for industrial scenarios. Human motion data created algorithmically lacks the variability and naturalness of real human motions; therefore, using motion-capture-based or human skeletal data extracted from RGB data is the preferred method for gathering human motion data for generation. However, current human motion generation models are trained on datasets that focus on activities of daily living, thereby lacking the terminology and body movements required in industrial environments. This work represents the first step towards identifying the requirements of human motion generation models. It provides a working example of industrial datasets for human motion generation with the instance of LARa.

Speaker: Nilah Nair

2:15 PM Grab your coffee and head to the next session

Transfer from first to ground floor

Collaborative Institutes/Initiatives: Introduction

Networking and growing together: The Lamarr Institute is embedded in and constantly growing its network. Get to know some of our partners that you can cooperate with through the Lamarr Institute.

67 Robotics Institute Germany (RGI)

The RIG comprises 14 German universities and research institutions and focuses on research clusters that cover all central areas of development and innovation in AI-powered robotics, enabling the translation of innovations into practice through close dialogue with industry.

Maren Bennewitz will present the RIG and its opportunities for collaboration.

Speaker: Maren Bennewitz

68 PhenoRob - Robotics and Phenotyping for Sustainable Crop Production

PhenoRob is one of six Clusters of Excellence at the University of Bonn and the only Cluster of Excellence in agriculture in Germany. It performs world-leading research in robotics and phenotyping for sustainable crop production. It hereby transforms crop production by optimizing breeding and farming management through developing and deploying new technologies.

Lamarr PI and PhenoRob spokesperson, Cyrill Stachniss, will present the Cluster of Excellence and its opportunities for collaboration.

Speaker: Prof. Cyrill Stachniss

69 Research Center Trustworthy Data Science and Security (RC Trust)

As part of the University Alliance Ruhr and founded in 2021, the RC Trust follows an interdisciplinary research approach that covers the entire spectrum of research challenges in all facets of trustworthy and privacy-aware technologies. Main foci include Artificial Intelligence and Machine Learning, Psychology and Social Sciences, Data Science and Statistical Learning as well as Cybersecurity and Privacy.

Lamarr Principal Investigator and Director of the RC Trust, Emmanuel Müller, will introduce the Research Center and its opportunities for collaboration.

Speaker: Emmanuel Müller

70 Bonn Sustainable AI Lab at IWE

As part of the The Institute for Science and Ethics (IWE), the Bonn Sustainable AI Lab postualtes sustainable AI as AI for sustainability and sustainability of AI. It aims to measure and assess the diverse environmental impacts of AI, research ways of making AI systems more sustainable, and address AI in the context of the Sustainable Development Goals.

The Bonn Sustainable AI Lab and its opportunities for collaboration will be presented.

Speaker: Dr Thomas Metcalf (University of Bonn)

71 Closing of the Event