Publication Server of Zuse Institute Berlin (ZIB)
Not a member yet
6648 research outputs found
Sort by
Warm-starting Strategies in Scalarization Methods for Multi-Objective Optimization
No full textWe explore how warm-starting strategies can be integrated into scalarization-based approaches for multi-objective optimization in (mixed) integer linear programming. Scalarization methods remain widely used classical techniques to compute Pareto-optimal solutions in applied settings. They are favored due to their algorithmic simplicity and broad applicability across continuous and integer programs with an arbitrary number of objectives. While warm-starting has been applied in this context before, a systematic methodology and analysis remain lacking. We address this gap by providing a theoretical characterization of warm-starting within scalarization methods, focusing on the sequencing of subproblems. However, optimizing the order of subproblems to maximize warm-start efficiency may conflict with alternative criteria, such as early identification of infeasible regions. We quantify these trade-offs through an extensive computational study
Fully automated quantification of synaptic locations in multi-channel Drosophila photoreceptor microscopy data
No full textThe workload posed by image analysis remains a major bottleneck for advances across the life sciences. To address this challenge, we have developed a fully automated workflow for processing complex 3D multi-channel microscopy images. Specifically, our workflow addresses the analysis of photoreceptor synapses in confocal images of the Drosophila melanogaster optic lobe. The workflow consists of multiple stages, combining traditional and machine learning–based approaches for image analysis and visual computing. It performs segmentation of brain regions, photoreceptor instance identification, and precise localization of synapses. The key novelty of the workflow is an automatic alignment of synapses into a cylindrical reference coordinate system, enabling comparative synaptic analysis across photoreceptors. To demonstrate the workflow’s applicability, preliminary biological results and their interpretation based on 50 images are presented. While the workflow is still being improved further, here, we showcase its capacity for efficient and objective data processing for high-throughput neurobiological analyses
Trust Region Constrained Measure Transport in Path Space for Stochastic Optimal Control and Inference
No full textSolving stochastic optimal control problems with quadratic control costs can be viewed as approximating a target path space measure, e.g. via gradient-based optimization. In practice, however, this optimization is challenging in particular if the target measure differs substantially from the prior. In this work, we therefore approach the problem by iteratively solving constrained problems incorporating trust regions that aim for approaching the target measure gradually in a systematic way. It turns out that this trust region based strategy can be understood as a geometric annealing from the prior to the target measure, where, however, the incorporated trust regions lead to a principled and educated way of choosing the time steps in the annealing path. We demonstrate in multiple optimal control applications that our novel method can improve performance significantly, including tasks in diffusion-based sampling, transition path sampling, and fine-tuning of diffusion models
Underdamped Diffusion Bridges with Applications to Sampling
No full textWe provide a general framework for learning diffusion bridges that transport prior to target distributions. It includes existing diffusion models for generative modeling, but also underdamped versions with degenerate diffusion matrices, where the noise only acts in certain dimensions. Extending previous findings, our framework allows to rigorously show that score matching in the underdamped case is indeed equivalent to maximizing a lower bound on the likelihood. Motivated by superior convergence properties and compatibility with sophisticated numerical integration schemes of underdamped stochastic processes, we propose \emph{underdamped diffusion bridges}, where a general density evolution is learned rather than prescribed by a fixed noising process. We apply our method to the challenging task of sampling from unnormalized densities without access to samples from the target distribution. Across a diverse range of sampling problems, our approach demonstrates state-of-the-art performance, notably outperforming alternative methods, while requiring significantly fewer discretization steps and no hyperparameter tuning
Universal and Expressive Statistical Shape Models for Anatomical Structures
No full textForm and function of anatomical structures are intimately linked. Pathological changes in form can be associated with the loss of function. For example, diseases often cause characteristic shape changes, making shape a sensitive structural biomarker for medical diagnosis. If the link between form and function is causal, correcting a pathological shape can even restore the healthy function of an organ. Accurate shape reconstruction is then crucial for effective, patient-specific treatment planning. This demonstrates the importance of shape in clinical interventions and its potential to improve overall patient outcomes.
Statistical shape models are computational methods that capture shape variations in a given population and enable precise shape analysis and generation. We focus on two key properties of a good statistical shape model. First, it should be easy to construct, and second, it should accurately represent the underlying shape distribution. Established existing approaches can only be constructed from surfaces with pre-defined dense correspondence. Such correspondence is tedious to obtain, can introduce undesired biases, and prevents training on partial or sparse observations. While correspondence-free methods exist, they struggle to accurately capture shape distributions with intricate details and large variations.
In this thesis, we develop shape models that simplify training and improve accuracy over state-of-the-art. To achieve these goals, we build on approximately diffeomorphic neural deformations and implicit neural representations. First, our proposed methods are trainable on correspondence-free surfaces and even partial segmentations with large slice distances. This makes them universal since they can be trained on heterogeneous data, enabling scalability to large datasets and avoiding potential biases of pre-defined correspondence. Second, our methods are highly expressive, accurately capturing intricate shape details in complex distributions. We evaluate effectiveness of our models on multiple anatomical structures, outperforming established baselines in both generative and discriminative settings
Column Generation for Periodic Timetabling
No full textPeriodic timetabling for public transportation networks is typically modelled as a Periodic Event Scheduling Problem (PESP). Solving instances of the benchmark library PESPlib to optimality continues to pose a challenge. As a further approach towards this goal, we remodel the problem by a time discretization of the underlying graph and consider arc-based as well as path-based integer programming formulations. For the path-based case, we use cycles on the graph expansion of the operational lines as variables and, therefore, include more of the problem inherent structure into the model. A consequence is the validity of several known inequalities and a lower bound on the LP-relaxation, that is the best known to date. As an extension we integrate passenger routing into the new model.
The proposed models have an advantage in the linear programming relaxation, on the one hand, but have an increased problem size, on the other hand. We define the corresponding pricing problems for the use of column generation to handle the size. Both models are practically tested on different problem instances
The Good, the Bad and the Ugly: Meta-Analysis of Watermarks, Transferable Attacks and Adversarial Defenses
No full text
Non-Reflecting Characteristic Boundary Conditions for Adjoint Time-Domain Acoustic Simulations
No full textAccurate acoustic simulations in the free field require non-reflective boundary conditions to suppress spurious reflections at the computational domain boundaries. Although several characteristic-based formulations for direct (forward) simulations have been proposed in recent decades, the adjoint formulations of such characteristic-based boundary conditions (CBCs) have received limited atten- tion in the literature and lack a comprehensive analysis. This paper presents the derivation and evaluation of adjoint CBCs complementing the existing direct CBCs. Both the forward and adjoint CBCs are applied to the (nonlinear) Euler equations and linear acoustic equations in time-domain simulations. In this manner, the CBCs are investigated and subsequently compared to assess their respective accuracy and consistency. The CBCs were implemented using both a single-point and a zonal approach, with the former optionally combined with a sponge layer. Both approaches yielded comparable results in direct and adjoint simulations, while the zonal CBC exhibited improved ac- curacy at lower frequencies. Across the evaluated frequency range of approximately 350 – 5600 Hz, spurious reflections were attenuated by up to −70dB in both forward and adjoint cases, demon- strating the effectiveness and consistency of the proposed boundary treatment
Combining LLMs and Knowledge Graphs to Reduce Hallucinations in Biomedical Question Answering
No full textAdvancements in natural language processing (NLP), particularly Large Language Models
(LLMs), have greatly improved how we access knowledge. However, in critical domains
like biomedicine, challenges like hallucinations—where language models generate infor-
mation not grounded in data—can lead to dangerous misinformation. This paper presents
a hybrid approach that combines LLMs with Knowledge Graphs (KGs) to improve the
accuracy and reliability of question-answering systems in the biomedical field. Our method,
implemented using the LangChain framework, includes a query-checking algorithm that
checks and, where possible, corrects LLM-generated Cypher queries, which are then exe-
cuted on the Knowledge Graph, grounding answers in the KG and reducing hallucinations
in the evaluated cases. We evaluated several LLMs, including several GPT models and
Llama 3.3:70b, on a custom benchmark dataset of 50 biomedical questions. GPT-4 Turbo
achieved 90% query accuracy, outperforming most other models. We also evaluated prompt
engineering, but found little statistically significant improvement compared to the standard
prompt, except for Llama 3:70b, which improved with few-shot prompting. To enhance
usability, we developed a web-based interface that allows users to input natural language
queries, view generated and corrected Cypher queries, and inspect results for accuracy. This
framework improves reliability and accessibility by accepting natural language questions
and returning verifiable answers directly from the knowledge graph, enabling inspection
and reproducibility. The source code for generating the results of this paper and for the user-
interface can be found in our Git repository: https://git.zib.de/lpusch/cyphergenkg-gui,
accessed on 1 November 2025