Autodidactic dense anatomical models.

Performance of large language models conducting systematic review tasks in prosthodontics.

BAED: A new paradigm for few-shot graph learning with explanation in the loop.

BCG-BP-FSNet: A few-shot learning framework for personalized, non-invasive blood pressure prediction based on ballistocardiogram

CaMST: Certainty-aware matching self-training for semi-supervised few-shot learning

Wind turbine gearboxes play a crucial role in wind energy production, and their failures can lead to significant downtime and maintenance costs. However, gearbox fault diagnosis remains a major challenge, especially with limited data. This study aims to develop a few-shot learning-based fault diagnosis model capable of accurately identifying wind turbine gearbox faults using a limited number of fault samples. Furthermore, this study also considers developing a remote diagnostic system for real-time monitoring and fault detection. The proposed framework combines data mining, feature extraction, and model optimization techniques with a remote monitoring system to achieve continuous fault detection. Experimental results show that even with limited data, the few-shot learning model outperforms traditional fault diagnosis methods, demonstrating higher accuracy and robustness. The remote diagnostic system also provides efficient real-time data transmission and fault detection, proving its practical value in continuous monitoring of wind turbine gearboxes. This research not only provides a practical solution for wind turbine gearbox fault diagnosis in data-scarce environments but also highlights the potential of remote diagnostics in real-time monitoring of wind turbine faults, paving the way for more efficient and economical maintenance strategies.

To address the evolving dynamics of network intrusion threats, this study proposes a Model-Agnostic Meta-learning (MAML)-based network attack detection model that utilizes an error-corrected few-shot dataset to enhance generalization across diverse attack scenarios. The model integrates an optimized Convolutional Neural Network (CNN) architecture with meta-learning algorithms, enabling adaptive feature extraction and rapid task-specific adaptation. By innovatively combining few-shot learning with CNN-based meta-learning, the framework achieves robust performance in detecting both simulated and real-world attacks. Empirical evaluations demonstrated that the model achieved near-optimal loss convergence (loss value close to 0) and exhibited an excellent training curve, indicating strong optimization stability. In real-world deployments, the model achieved a defense action probability of 0.85 for simulated attacks, a detection accuracy of 0.91, and a recognition accuracy of 0.92 for real-world attacks, outperforming baseline methods in dynamic network environments. These results highlight the model's effectiveness in improving network platform defenses against complex and evolving attack behaviors, thereby advancing the most advanced in adaptive intrusion detection systems.

Spiking Neural Networks (SNNs) have emerged as a promising paradigm for brain-inspired edge computing. Leveraging binary spikes and local learning rules, SNNs enable energy-efficient on-chip learning and rapid adaptation to changing environments, which is crucial for edge AI that needs to learn continuously from new data. However, many SNN processors enabling on-chip learning for edge computing confront a trade-off: small-scale task-specific designs offer low power but poor multi-task inference accuracy, while large-scale general-purpose designs achieve high multi-task accuracy at the cost of large memory and poor energy efficiency. To overcome this challenge, this paper presents ANP-R, a 22nm asynchronous SNN-based edge AI processor with coarse-grained reconfigurable architecture enabling one-shot, few-shot, batch and incremental on-chip learning. The processor integrates 64 cores containing 4096 neurons and 0.262 million synapses. Two key features are proposed: 1) An asynchronous coarse-grained reconfigurable architecture that supports various STDP-based SNN topologies. These topologies enable over 95% average accuracy across four sensory tasks; 2) an energy-efficient asynchronous training method incorporating a self-adaptive synaptic weight update mechanism reducing up to 65% redundant updates without accuracy loss, and a trained weights low-bit width coding method reducing up to 50% storage cost with 0.3% accuracy loss. Measurement results demonstrate 92.1% accuracy for hand gesture classification, 93.9% for keyword spotting, 98.6% for object recognition and 99.2% for gas identification. Compared with state-of-the-art SNN-based chips, this work achieves up to 6.02x, 8.61x and 7.1% improvement in energy efficiency, energy per step, and accuracy, respectively.

Designing enzyme sequences to enhance product yield represents a fundamental challenge in metabolic engineering. Here, we established a workflow that integrates computational predictions with efficient experimental iteration to obtain outsized gains in product yield. Based on causal inference and examination of published datasets, we realized and ultimately experimentally confirmed that in vivo unit yield (yield/expression) can serve as an attractive surrogate for aqueous kcat/Km when optimizing for activity. In our workflow, we initially predict activity-enhancing single mutants by calculating the binding affinities of reactive intermediates, followed by experimental investigations of unit yield. Subsequently, we predict activity-enhancing mutation combinations using a few-shot learning model we developed called Physics-Inspired Feature Selection of Protein Language Models (PIFS-PLM), which requires only 60-100 experimentally examined mutation combinations as input. In a case study of a bicyclogermacrene (BCG) synthase, we achieve a 73-fold increase in BCG yield or a 15% increase in BCG selectivity based on combinations of 12 individual mutations, and provide extensive crystallographic and biochemical evidence for impacts from specific mutations. Thus, optimizing for unit yield is highly efficient as an alternative to optimizing for thermostability, and our study provides a powerful workflow for the efficient engineering of high-yield enzyme variants.

Falls pose a significant threat to human safety, making rapid and accurate detection and response essential. Time Exploration Network (TExNet), an attention-enhanced network tailored for identifying falls accurately and rapidly, is proposed in this paper. Unlike existing studies that rely on simulated environments, TExNet addresses the gap between simulated and real scenarios by integrating multi-branch timing and classification characteristics. It features a two-branch adaptive fusion framework, leveraging Convoluational Neural Network (CNN) and Transformer architectures to capture both local and global dependencies effectively. Additionally, dual-branch adaptive fusion framework incorporates dilated convolution and time series positional decomposition to enhance temporal correlation understanding. To handle data distribution variations, it employs an Invariant Risk Minimization (IRM) inspired loss function, penalizing misclassification of positive examples. This reduces model reliance on specific environments and improves action understanding. Moreover, a data self-conditioning module enhances data diversity and tackles imbalance issues. The model is deployed using a fine-tuning strategy based on a pre-trained framework combined with few-shot learning for downstream tasks. Experimental results show that the fine-tuned model achieves a recall of 92.16%, indicating its strong ability to rapidly adapt to new data distributions. Extensive experiments also validate the superiority of TExNet compared with existing approaches.

Generative information extraction using large language models (LLMs), particularly through prompting combined with few-shot learning, has become a popular method. In many ways such prompts with examples resemble the annotation guidelines long used for manual labeling of data for information extraction, and indeed studies have demonstrated the direct use of these guidelines as effective prompts. However, constructing annotation guidelines is both labor- and knowledge-intensive. Instead, this paper proposes to leverage LLMs' impressive ability to automatically create such annotation guidelines. Specifically, we propose a zero-shot hierarchical prompt engineering method that harvests the knowledge summarization and text generation capacity of LLMs to synthesize annotation guidelines to improve downstream LLMs while requiring minimal human input. Zero-shot clinical named entity recognition benchmarks, 2012 i2b2 EVENT, 2012 i2b2 TIMEX, 2014 i2b2, and 2018 n2c2 showed improvements of 0.2% to 25.86% for Llama 3.1 and 5.82% to 16.13% for GPT-OSS in strict F1 scores from the no-guideline baseline. The LLM-synthesized guidelines showed equivalent or better performance compared to human-written guidelines by 0.23% to 10.00% in most tasks. LLMs generate high-quality annotation guidelines following a consistent pattern (eg, title, entity types, examples) without human guidance, indicating that a representation of such a concept has been encoded during the pre-training. Nuances in definitions, however, still require adjustment by researchers to align with the project. This study proposes a novel hierarchical prompt engineering method that requires minimal knowledge transfer from a human expert and is applicable to multiple biomedical domains.

Artificial intelligence and organic chemistry are redefining each other in a fundamentally bidirectional relationship. This Review highlights how the intrinsic challenges of organic chemistry have acted as a catalyst for conceptual and methodological innovation in AI itself. Sparse and heterogeneous reaction data sets spurred the development of self-supervised and few-shot learning paradigms; the combinatorial complexity of multireactant chemistry motivated the transition from graph neural networks to hypergraph architectures; the need to bridge symbolic chemical reasoning with statistical prediction inspired chemical language models grounded in large language model frameworks; and the iterative, decision-intensive nature of synthesis planning catalyzed the rise of autonomous agentic systems. We survey the multimodal landscape of chemical data, tracing the evolution of molecular representations from classical fingerprints to geometric encodings and examining how each representation class shapes downstream model capabilities. We analyze how data scarcity and uneven property distributions have driven advances in transfer learning, self-supervised pretraining, and meta-learning frameworks tailored to molecules and reactions. Reaction prediction, mechanistic inference, and retrosynthesis planning are examined as core areas where chemistry has shaped modern AI techniques. We further explore chemical reasoning through multimodal fusion, generative molecular design, and self-driving laboratories. We conclude by identifying persistent challenges, including data sparsity, selection bias, benchmark-to-lab gaps, and reproducibility.

Safety supervision at power operation sites is critical for ensuring worker safety and maintaining a reliable electricity supply. However, existing safety violation detection methods are constrained by limited labeled data, poor performance on small-object detection tasks, and interference from complex backgrounds. To overcome these challenges, this study proposes a framework that integrates multi-scale object detection with few-shot learning. A multi-scale feature extraction module is designed based on a feature pyramid network and channel attention mechanisms to enhance the perception of small objects. In addition, a few-shot learning framework incorporating a meta-learning strategy is introduced to address the scarcity of labeled safety violation samples and improve the model's adaptability to new tasks with limited training data. Experimental results demonstrate that the proposed method consistently outperforms existing approaches across multiple evaluation metrics. The framework achieves notable improvements in small-object detection accuracy and few-shot learning performance, resulting in enhanced detection accuracy, robustness, and generalization capability. The integration of multi-scale feature extraction and few-shot learning effectively addresses the challenges of safety violation detection in power operation environments. The proposed framework provides a practical and reliable solution for intelligent safety monitoring and has significant potential for real-world deployment in power operation sites.

Machine vision-based surface defect detection offers the advantages of being non-contact, non-destructive, and highly automated; consequently, it is widely applied across various industrial production processes. This article provides a brief overview of commonly used methods for surface defect detection, evaluation indicators for detection results, and key challenges currently faced. Defect detection methods are categorized into three types: traditional image processing methods, traditional machine learning methods, and deep learning methods. The core principles and representative studies of each method are reviewed, and their respective advantages and limitations are analyzed. We briefly describe the method for evaluating detection results, examine the few-shot learning problem encountered in practical applications, and provide an outlook on feasible pathways for addressing this issue in the future.

Objective: To develop a micro-CT pulp cavity image segmentation model based on few-shot transfer learning, enabling efficient and accurate segmentation of the pulp cavity with limited training samples, thereby supporting three-dimensional(3D) anatomical research of the pulp cavity and digital root canal treatment. Methods: Extracted teeth (n=110) due to pathological reasons were collected from the Department of Oral and Maxillofacial Surgery, Hospital of Stomatology, Wuhan University, between January 2025 and September 2025. These teeth were scanned using micro-CT. The acquired images were randomly divided (simple random sampling) into a training set (10 teeth), a test set (90 teeth), and an independent test set (10 teeth) at a ratio of 1∶9∶1. 3D pulp cavity images were meticulously annotated by two oral clinicians (attending physicians) to establish the ground truth for segmentation. By introducing a cross-domain adaptation strategy, the natural image segmentation foundation model, segment anything model (SAM), was transferred to the task of pulp cavity image segmentation under the guidance of a limited number of training samples, thereby constructing an automatic micro-CT pulp cavity image segmentation model (PulpSAM). Multi-dimensional quantitative and qualitative analyses were performed to compare the segmentation performance on the pulp cavity (entire pulp cavity, apical 3 mm, and lateral accessory canals) achieved by deep learning models obtained with different numbers of training samples (0, 1, 3, and 10 teeth), i.e., SAM, PulpSAM-1, PulpSAM-3, and PulpSAM-10. These results were also compared with those of the U-Net and nnU-Net models. Additionally, the time required for three methods (manual annotation, fully automated model segmentation, and model segmentation followed by manual refinement) was compared. Results: Under training sample sizes of 1, 3, and 10 teeth, all PulpSAM models achieved high segmentation performance, with median Dice coefficient, Intersection over Union (IoU), precision, recall, and accuracy all≥92.3%, median 95% Hausdorff distance (95%HD)≤0.04 mm, and median average symmetric surface distance (ASSD) of 0.02 mm. These metrics were significantly superior to those of the SAM model (all P<0.008 3). The pulp cavity volume and relative volume difference (RVD) segmented by each PulpSAM model were significantly smaller than those of the SAM model (all P<0.008 3). Except for ASSD and precision, the other seven segmentation accuracy metrics of PulpSAM-3 were significantly better than those of PulpSAM-1 (all P<0.008 3). Compared with PulpSAM-10, PulpSAM-3 showed significantly higher recall, ASSD, and pulp cavity volume (all P<0.008 3), and significantly lower precision and absolute RVD (all P<0.008 3). No statistically significant differences were observed between PulpSAM-3 and PulpSAM-10 in Dice coefficient, IoU, accuracy, or 95%HD (all P>0.008 3). All segmentation accuracy metrics of each PulpSAM model were significantly superior to those of the corresponding U-Net and nnU-Net models (all P<0.017). The segmentation times required for manual annotation, PulpSAM-3 automatic segmentation, and PulpSAM-3 segmentation followed by manual refinement were 3 354.6 (852.0), 190.0 (27.9), 646.4 (171.5) s, respectively, with statistically significant differences among the three methods (χ²=25.80, P<0.001). Conclusions: This study developed the PulpSAM model based on a few-shot transfer learning strategy, achieving efficient and accurate automatic segmentation of the pulp cavity in micro-CT images.

Medical image analysis is essential for modern diagnostics, as it enables accurate and rapid disease detection. However, traditional deep learning models require large, annotated datasets, which are frequently inaccessible in medical scenarios due to data scarcity, privacy constraints, and excessive labeling costs. Few-Shot Learning (FSL) and Automated Machine Learning (AutoML) have appeared as effective techniques to address these issues. FSL utilizes meta-learning and metric-based strategies enabling models to learn from small samples, while AutoML automates model design and optimization, reducing reliance on expert intervention. Additionally, the integration of domain-specific knowledge, including anatomical priors and clinically relevant features, has demonstrated enhancements in interpretability and diagnostic significance. This mini review offers a structured analysis of FSL, AutoML, and domain-specific knowledge in medical image analysis, emphasizing their potential integration. A critical evaluation of existing literature reveals that most studies use these approaches independently. The review further examines methodological limitations, dataset constraints, and clinical applicability challenges across current studies. Based on these findings, key research gaps are identified, such as the need for domain-informed architecture search, standardized evaluation protocols, and pipelines that use less computer power. Notably, metric-based FSL approaches were more widely used than gradient-based methods due to their stability under limited data conditions. However, the literature is still methodologically fragmented, with FSL, AutoML, and domain-specific knowledge mainly studied separately or in partial combinations. The paper concludes by outlining future research directions toward the development of AutoML-enhanced FSL frameworks integrated with domain-specific knowledge to improve performance, interpretability, and clinical reliability in data-constrained environments.

Accurate identification of protein particles in cryo-electron microscopy (cryo-EM) micrographs is crucial for high-resolution structure determination, but remains challenging due to the heavy reliance on extensive annotated datasets and the difficulty of ensuring robustness under low signal-to-noise ratio (SNR) conditions. Current approaches require large annotations and exhibit poor generalization to new protein targets. We present CryoFSL (Cryo-EM Few Shot-Learning), a novel few-shot learning framework built on Segment Anything Model 2 with lightweight adapters, enabling robust particle picking with as few as five labeled micrographs and significantly reducing the annotation burden. The framework’s hierarchical adapter design supports dynamic feature modulation for low-SNR and heterogeneous conditions, resolving the trade-off between annotation burden and performance. CryoFSL surpasses both traditional template-based methods and state-of-the-art deep learning models across diverse proteins in the few-shot learning setting, achieving superior recall, precision, and 3D reconstruction resolution with minimal supervision. It maintains stability across heterogeneous micrographs and consistently detects high-quality particles with fewer false-positives. Notably, CryoFSL achieves competitive resolution in density map reconstruction with just a fraction of the particles picked by other methods, redefining efficiency and quality in cryo-EM analysis. This work paves the way for scalable, generalizable, and annotation-efficient particle-picking pipelines. The code is available at https://github.com/biplabpoudel25/CryoFSL.

Few-shot learning photovoltaic fault diagnosis based on Tri-MSTCN

Deep learning (DL) methods have demonstrated promising performance in online motor unit (MU) identification from high-density surface electromyogram (HD-sEMG). However, its dependency on larger amounts of data limits practicability. To address this issue, a novel few-shot learning method enhanced by sample reconstruction strategy is presented for online MU identification. In this method, a spatio-temporal neural network was first pre-trained based on the simulated HD-sEMG data and MU spike trains, endowing it with initial representational capabilities for MU features. Then, an innovative sample reconstruction strategy was employed to generate physiologically interpretable synthetic samples for model fine-tuning with minimal experimental data. These samples were intended to enhance the model's ability to characterize the spatiotemporal features of MUAP waveforms, thereby improving the online MU identification performance. Experimental HD-sEMG signals were collected from the abductor pollicis brevis muscles of ten subjects using an 8 × 8 electrode array. The results demonstrated that the proposed method can achieve a matching rate of approximately 93% in online MU identification stage, with only 6s of experimental data for offline model fine-tuning, significantly outperforming the comparison methods. This work provides a novel solution for efficient real-time MU identification, and the findings are expected to advance the widespread application of DL-based real-time HD-sEMG decomposition in developing advanced neural-machine interfaces towards robotic motor control and rehabilitation medicine.

ABSTRACT Currently, a variety of land surface temperature (LST) products generated from thermal infrared bands have been already accumulated. Compared to the thermal infrared band, the mid-infrared band exhibits higher transmittance and greater robustness under humid atmospheric conditions, offering potential for further improving LST retrieval accuracy. However, the mid-infrared band presents larger variability in emissivity and higher estimation difficulty, limiting the effectiveness of LST retrieval using mid-infrared remote sensing data sources. This study proposed a MODIS night-time mid-infrared LST retrieval algorithm that integrates reflectance spectral characteristics to estimate emissivity. A few-shot machine learning model was established to build the correlation between MODIS optical band reflectance and mid-infrared band emissivity within a simulated dataset accounting for mixed spectral components, then applied to real observational data. Validation results from SURFRAD ground stations indicate an overall RMSE of 2.5521 K for this new algorithm, with values of 2.5558 K under dry atmospheric conditions and 2.5021 K under humid atmospheric conditions. The new algorithm can accurately retrieve night-time LST without significant error increasing as atmospheric water vapor content rises. Future work will further study on fields including eliminating daytime solar radiance effects, conducting multi-surface-type validation, and reducing dependence on external parameters.