Patterns最新文献

Recent studies have demonstrated that high-quality annotated data are crucial for segmentation performance. However, incomplete or corrupted mask annotations remain common, limiting supervised learning. To address this, we introduce a mask-reconstruction task, referred to as masked segmentation label modeling (MSLM), which refines partially occluded labels by leveraging visible regions without manual annotations. We further propose the label masked autoencoder (L-MAE), which identifies erroneous regions and reconstructs them through contextual inference. The L-MAE fuses incomplete labels and corresponding images into a unified map for reconstruction, and an image patch supplement (IPS) algorithm restores missing image information, improving the average mean intersection over union (mIoU) by 4.1%. To validate the L-MAE, we train segmentation models on a degraded and L-MAE-enhanced Pascal VOC dataset, with the latter achieving a 13.5% mIoU improvement. The L-MAE attains predict area (PA)-mIoU scores of 91.0% on Pascal VOC 2012 and 86.4% on Cityscapes, outperforming state-of-the-art supervised segmentation models.

Autonomous AI-to-AI creative systems promise new frontiers in machine creativity, yet we show that they systematically converge toward generic outputs. We built iterative feedback loops between Stable Diffusion XL (SDXL; image generation) and Large Language and Vision Assistant (LLaVA; image description), forming autonomous text → image → text → image cycles. Across 700 trajectories with diverse prompts and 7 temperature settings over 100 iterations, all runs converged to nearly identical visuals-what we term "visual elevator music." Quantitative analysis revealed just 12 dominant motifs with commercially safe aesthetics, such as stormy lighthouses and palatial interiors. This convergence persisted across model pairs, indicating structural limits in cross-modal AI creativity. The effect mirrors human cultural transmission, where iterated learning amplifies cognitive biases, but here, diversity collapses entirely as AI loops gravitate to high-probability attractors in training data. Our findings expose hidden homogenizing tendencies in current architectures and underscore the need for anti-convergence mechanisms and sustained human-AI interplay to preserve creative diversity.

Although there are ways to estimate the global power demand of artificial intelligence (AI) systems, it remains challenging to quantify the associated carbon and water footprints. The lack of distinction between AI and non-AI workloads in the environmental reports of data center operators makes it possible to assess the environmental impact of AI workloads only by approximating them through data centers' general performance metrics. The environmental disclosure of tech companies is, however, often insufficient to determine even the total data center performance of these companies. The shortcomings in the environmental disclosure of data center operators could be remedied with new policies mandating the disclosure of additional metrics. Because the environmental impact of data centers is growing rapidly, the urgency of transparency in the tech sector is also increasing. The carbon footprint of AI systems alone could be between 32.6 and 79.7 million tons of CO₂ emissions in 2025, while the water footprint could reach 312.5-764.6 billion L.

Record linkage and pseudonymization are crucial tasks in collaborative biomedical research. Data for a patient are rarely stored in one place and therefore often need to be linked and integrated across multiple institutions. Mainzelliste is an open-source software solution designed to solve these challenges by providing a comprehensive and flexible toolkit for pseudonymization, record linkage, and consent management. It supports a variety of pseudonyms, record linkage methods, and modular, informed patient consents. A highly flexible REST application programming interface (API) allows tight integration into existing applications and workflows. Since its initial release in 2015, Mainzelliste has evolved into a vibrant open-source software solution "by researchers, for researchers" including a user-friendly graphical interface, support for HL7 FHIR for consent and patient data, and record linkage based on secure multi-party computation, thereby supporting secure and efficient biomedical research.

Sleep is one of the most essential parts of our daily lives. The gold standard for studying sleep is polysomnography (PSG) recordings. The first step of analyzing PSG recordings involves splitting them into sleep stages, which is performed manually. Machine learning algorithms have attempted to automate the tedious task of sleep scoring, mostly via supervised learning. A recent study in Patterns introduces AISleep, a novel algorithm approaching the task of sleep scoring in an unsupervised framework. This algorithm is based on humanly interpretable features and provides robust results across different datasets and age groups.

Standard offline evaluations for language models fail to capture how these models actually behave in practice, where personalization fundamentally alters model behavior. In this work, we provide empirical evidence showcasing this phenomenon by comparing offline evaluations to field evaluations conducted by having 800 real users of ChatGPT and Gemini pose benchmark and other questions to their chat interfaces.

Emphysema, a diffuse and heterogeneous phenotype of chronic obstructive pulmonary disease (COPD), carries substantial morbidity and elevates lung cancer risk. While computed tomography (CT) aids in detection and monitoring, current deep learning methods depend on large annotated datasets. Unsupervised anomaly detection (UAD) provides an alternative but faces challenges with emphysema anomalies and weak emphysema semantics. In this study, we propose a self-supervised framework trained exclusively on non-emphysema CT scans using synthetically generated lesions to guide pixel-level anomaly modeling. We introduce EDLNet, an encoder-decoder architecture with spatial-channel refinement and adaptive feature fusion for emphysema detection and localization, followed by an unsupervised manner for emphysema staging. Multi-center evaluations show that our framework outperforms existing UAD approaches in detection and localization, while achieving a mean staging accuracy of 93.13% and a macro AUROC of 99.08%. This approach bridges clinical knowledge and artificial intelligence, offering a scalable and interpretable solution for lung disease analysis.

DNA barcode networks are the basis of sequencing-based microscopy, an emerging family of chemical imaging methods aiming to reconstruct spatial information, without optics, using sequencing technology. These methods capture microscopic spatial information by forming networks composed of many local chemical interactions, each marked by a unique, DNA-based barcode. However, the fundamental laws governing such networks are not yet understood, and spatial barcode networks are influenced by structural distortions such as false or shortcut edges. Current methods lack ground-truth-free tools to validate spatial quality, and we address this with a framework for topology-based quality control. We define a fundamental feature of spatial networks, spatial coherence, which quantifies geometric self-consistency in a network. By formalizing this relationship into quantitative metrics adapted from classical geometric rules, we could quantify spatial distortions by using only network data and show how these can be used as an optimization criterion to iteratively improve spatial reconstruction.

Protein language models (PLMs) are increasingly used to assess the impact of genetic variants, achieving high accuracy and often outperforming traditional pathogenicity predictors. They enable zero-shot inference, making predictions without task-specific fine-tuning, though studying in-frame insertions and deletions (indels) remains challenging due to altered protein lengths and limited annotated datasets. Here, we present IndeLLM, a scoring approach for indel pathogenicity that accounts for sequence length differences. Our zero-shot method relies solely on sequence information, requires minimal computing resources, and achieves performance comparable to existing predictors. Building on this, we developed a Siamese network via transfer learning that outperformed all tested indel predictors (Matthews correlation coefficient = 0.77). To enhance accessibility, we provide a plug-and-play Google Colab notebook for using IndeLLM and visualizing the impact of indels on protein sequence and structure. The tool is freely available on GitHub and Google Colab.

Large language models (LLMs) can surpass humans in certain forecasting tasks. What role does this leave for humans in the overall decision process? One possibility is that humans, despite performing worse than LLMs, can still add value when teamed with them. A human and machine team can surpass each individual teammate when team members' confidence is well calibrated and team members diverge in which tasks they find difficult (i.e., calibration and diversity are needed). We simplified and extended a Bayesian approach to combining judgments using a logistic regression framework that integrates confidence-weighted judgments for any number of team members. Using this straightforward method, we demonstrated its effectiveness in both image classification and neuroscience forecasting tasks. Combining human judgments with one or more machines consistently improved overall team performance. Our hope is that this simple and effective strategy for integrating the judgments of humans and machines will lead to productive collaborations.