28 February 2025 in Journal of Computational Social Science
Evaluating bias in Large Language Models (LLMs) has become a pivotal issue in current Artificial Intelligence (AI) research due to their significant impact on societal dynamics. Recognizing political bias in LLMs is particularly important as they approach performative prediction, influencing societal behavior and political events, such as the upcoming European Parliament elections. From a German voter’s perspective, we evaluate the political bias of the currently most popular open-source LLMs concerning political issues within the European Union. To do so, we use the "Wahl-O-Mat," a voting advice application used in Germany. We show that larger models, such as Llama3-70B, tend to align more closely with left-leaning political parties, while smaller models often remain neutral, particularly when prompted in English. The central finding is that LLMs are similarly biased, with low variances in the alignment concerning a specific party. Our findings offer crucial insights for developers and policymakers to understand and mitigate LLM biases, emphasizing the need for rigorous bias assessment to ensure the integrity and trustworthiness of AI applications.
12 June 2024 in npj Computational Materials
Electron microscopy is indispensable for examining the morphology and composition of solid materials at the sub-micron scale. To study the powder samples that are widely used in materials development, scanning electron microscopes (SEMs) are increasingly used at the laboratory scale to generate large datasets with hundreds of images. Parsing these images to identify distinct particles and determine their morphology requires careful analysis, and automating this process remains challenging. In this work, we enhance the Mask R-CNN architecture to develop a method for automated segmentation of particles in SEM images. We address several challenges inherent to measurements, such as image blur and particle agglomeration. Moreover, our method accounts for prediction uncertainty when such issues prevent accurate segmentation of a particle. Recognizing that disparate length scales are often present in large datasets, we use this framework to create two models that are separately trained to handle images obtained at low or high magnification. By testing these models on a variety of inorganic samples, our approach to particle segmentation surpasses an established automated segmentation method and yields comparable results to the predictions of three domain experts, revealing comparable accuracy while requiring a fraction of the time. These findings highlight the potential of deep learning in advancing autonomous workflows for materials characterization.
6 March 2023 in IEEE Transactions on Biomedical Engineering
Objective : The scarcity of high-quality annotated data is omnipresent in machine learning. Especially in biomedical segmentation applications, experts need to spend a lot of their time into annotating due to the complexity. Hence, methods to reduce such efforts are desired. Methods : Self-Supervised Learning (SSL) is an emerging field that increases performance when unannotated data is present. However, profound studies regarding segmentation tasks and small datasets are still absent. A comprehensive qualitative and quantitative evaluation is conducted, examining SSL's applicability with a focus on biomedical imaging. We consider various metrics and introduce multiple novel application-specific measures. All metrics and state-of-the-art methods are provided in a directly applicable software package ( https://osf.io/gu2t8/ ). Results : We show that SSL can lead to performance improvements of up to 10%, which is especially notable for methods designed for segmentation tasks. Conclusion : SSL is a sensible approach to data-efficient learning, especially for biomedical applications, where generating annotations requires much effort. Additionally, our extensive evaluation pipeline is vital since there are significant differences between the various approaches. Significance : We provide biomedical practitioners with an overview of innovative data-efficient solutions and a novel toolbox for their own application of new approaches. Our pipeline for analyzing SSL methods is provided as a ready-to-use software package.
1 September 2023 in Current Directions in Biomedical Engineering
U-Net is the go-to approach for biomedical segmentation applications. However, it is not designed to segment overlapping objects, a challenge Mask R-CNN has shown to have great potential in. Yet, Mask R-CNN receives little attention in biomedicine. Hence, we evaluate both approaches on a publicly available biomedical dataset. We find that Mask RCNN outperforms U-Net in segmenting overlapping cells and achieves comparable performance if they do not intersect. Our study provides valuable decision support to practitioners in selecting an appropriate method when solving instance segmentation tasks using deep learning, as well as important insights into enhancing the accuracy of such approaches in biomedical image analysis.
8 February 2022 in PLOS ONE
Deep learning increasingly accelerates biomedical research, deploying neural networks for multiple tasks, such as image classification, object detection, and semantic segmentation. However, neural networks are commonly trained supervised on large-scale, labeled datasets. These prerequisites raise issues in biomedical image recognition, as datasets are generally small-scale, challenging to obtain, expensive to label, and frequently heterogeneously labeled. Furthermore, heterogeneous labels are a challenge for supervised methods. If not all classes are labeled for an individual sample, supervised deep learning approaches can only learn on a subset of the dataset with common labels for each individual sample; consequently, biomedical image recognition engineers need to be frugal concerning their label and ground truth requirements. This paper discusses the effects of frugal labeling and proposes to train neural networks for multi-class semantic segmentation on heterogeneously labeled data based on a novel objective function. The objective function combines a class asymmetric loss with the Dice loss. The approach is demonstrated for training on the sparse ground truth of a heterogeneous labeled dataset, training within a transfer learning setting, and the use-case of merging multiple heterogeneously labeled datasets. For this purpose, a biomedical small-scale, multi-class semantic segmentation dataset is utilized. The heartSeg dataset is based on the medaka fish’s position as a cardiac model system. Automating image recognition and semantic segmentation enables high-throughput experiments and is essential for biomedical research. Our approach and analysis show competitive results in supervised training regimes and encourage frugal labeling within biomedical image recognition.
22 September 2025 in npj Computational Materials
Scanning Electron Microscopes (SEMs) are widely used in experimental science laboratories, often requiring cumbersome and repetitive user analysis. Automating SEM image analysis processes is highly desirable to address this challenge. In particle sample analysis, Machine Learning (ML) has emerged as the most effective approach for particle segmentation. However, the time-intensive process of manually annotating thousands of SEM images limits the applicability of supervised learning approaches. Self-Supervised Learning (SSL) offers a promising alternative by enabling knowledge extraction from raw, unlabeled data. This study presents a framework for evaluating SSL techniques in SEM image analysis, focusing on novel methods leveraging the ConvNeXtV2 architecture for particle detection. A dataset comprising 25,000 SEM images is curated to benchmark these proposed SSL methods. The results demonstrate that ConvNeXtV2 models, with varying parameter counts, consistently outperform other techniques in particle detection across different length scales, achieving up to a 34% reduction in relative error compared to established SSL methods. Furthermore, an ablation study explores the relationship between dataset size and SSL performance, providing actionable insights for practitioners regarding model selection and resource efficiency. This research advances the integration of SSL into autonomous analysis pipelines and supports its application in accelerating materials science discovery.