Humpback whales produce a wide variety of frequency-modulated vocalizations, called song units. Modeling and synthesis of these units form the basis for many bio-inspired applications, including underwater covert communication and naturalistic playback experiments. Conventional synthesis methods are based on fundamental frequency contour modeling of single-segment signals, which exhibit limitations in terms of synthesis flexibility and similarity. To address the above limitations, this paper proposes a humpback whale song unit synthesis method based on small-sample training. Fundamental frequency contours and line spectral pairs are extracted from humpback whale song units collected in marine environments to construct the training dataset. Using these parameters, a hidden Markov model (HMM) is established for parameter training, and probability density functions are obtained for each HMM state. To address high-frequency jitter in generated fundamental frequency contours, a parameter generation method that combines dynamic feature constraints with variational mode decomposition denoising is introduced, yielding smoother fundamental frequency curves. For enhanced synthesis flexibility, state duration modification and fundamental frequency modification methods are proposed based on parameter distributions. Finally, the generated parameters are converted into time-domain waveforms using a linear predictive coding-pitch vocoder. To comprehensively evaluate the synthesis performance, an assessment framework based on statistical parametric analysis and t-distributed stochastic neighbor embedding is established. Simulation results demonstrate that the proposed humpback whale song unit synthesis system achieves superior flexibility and similarity compared to the conventional approach based on single whistles modeling, ultimately enhancing performance in bio-inspired applications.

Tokamak is a typical nuclear fusion device, where effective control of the plasma boundary contributes to its long-pulse stable operation. Nowadays, CCD camera diagnostic systems are widely used in most fusion devices. By detecting the CCD images of plasma discharge, the plasma boundary can be obtained. The U-Net deep learning network, with its excellent feature extraction and boundary perception capabilities, possesses excellent image detection characteristics and can be applied in boundary detection. However, automated and accurate detection of plasma boundaries from these images remains challenging due to their inherent complex characteristics, such as low contrast, blurry edges, nonuniform brightness distribution, and high inter-class similarity. To address the challenge of automated, high-precision plasma boundary detection in Tokamak devices, this paper develops a novel diagnostic method based on an optimized deep learning network. Firstly, batch normalization (BN) layers are introduced between each convolution operation and activation function in the encoder-decoder, effectively mitigating internal covariate shift during small-sample training and significantly improving training speed and gradient stability. Furthermore, by embedding the convolutional block attention module (CBAM) before each downsampling, the U-Net network is enabled to adaptively enhance the response of discriminative feature channels and key spatial regions, effectively distinguishing visually similar areas such as the plasma core, edge, and background. Additionally, a hybrid loss function combining cross-entropy and Dice loss is designed to enhance boundary sensitivity while maintaining smooth and stable training, thereby alleviating model bias caused by class imbalance. To verify the effectiveness of the detection algorithm, plasma discharge images through the CCD camera on the Thailand Tokamak-1 (TT-1) device are obtained, and a specialized dataset for plasma image detection is constructed. Experimental results demonstrate that the optimized model achieves a final mIoU of 92.42%, mPrecision of 96.67%, and mRecall of 95.24%, significantly outperforming both the original U-Net network and Canny edge detection. The proposed algorithm not only reduces misjudgment and boundary discontinuities but also exhibits stronger robustness and generalization capability, confirming its suitability for accurate plasma boundary detection.

Empirical simulation of internal validation methods for prediction models: comparing k-fold cross-validation with bootstrap-based optimism correction.

Sitting posture recognition, defined as automatically localizing and categorizing seated human postures, has become essential for large-scale ergonomics assessment and longitudinal health-risk monitoring in classrooms and offices. However, in real-world multi-person scenes, pervasive occlusions and overlaps induce keypoint misalignment, causing global-attention backbones to fail to localize critical local structures. Moreover, annotation scarcity makes small-sample training commonplace, leaving models insufficiently robust to misalignment perturbations and thereby limiting cross-domain generalization. To address these challenges, we propose LAViTSPose, a lightweight cascaded framework for sitting posture recognition. Concretely, a YOLOR-based detector trained with a Range-aware IoU (RaIoU) loss yields tight person crops under partial visibility; ESBody suppresses cross-person leakage and estimates occlusion/head-orientation cues; a compact ViT head (MLiT) with Spatial Displacement Contact (SDC) and a learnable temperature (LT) mechanism performs skeleton-only classification with a local structural-consistency regularizer. From an information-theoretic perspective, our design enhances discriminative feature compactness and reduces structural entropy under occlusion and annotation scarcity. We conducted a systematic evaluation on the USSP dataset, and the results show that LAViTSPose outperforms existing methods on both sitting posture classification and face-orientation recognition while meeting real-time inference requirements.

Medical image segmentation is fundamental for disease diagnosis, particularly in the context of breast cancer, a prevalent malignancy affecting women. The accuracy of lesion localization and preservation of image details are essential for ensuring the integrity of lesion segmentation. However, the low resolution of breast tumor B-mode ultrasound images poses challenges in precisely identifying lesion sites. To address this issue, this study introduces the Mamba architecture model, which combines three foundational models with the long-sequence processing model Mamba to develop a novel segmentation model for breast tumor ultrasound images. The selective mechanism and hardware-aware algorithm of the Mamba model enable longer sequence inputs and faster computing speeds. Moreover, integrating a complete chain of VMamba blocks into the basic model enhances segmentation accuracy and image detail processing capabilities. Experimental segmentation was performed on two benchmark ultrasound datasets (BUSI and BUS-BRA) using both the baseline and improved models. The results were compared using metrics such as Dice and IoU, with additional evaluations conducted under small-sample training conditions. This study is intended to provide guidance for the future development of medical image segmentation. Moreover, the experimental results demonstrate that the model incorporating the Mamba architecture achieves superior performance on breast ultrasound images.

Human locomotion mode recognition plays a crucial role in the effective operation of lower limb exoskeletons. This paper proposes a method that integrates Empirical Mode Decomposition (EMD) with data enhancement and bidirectional temporal convolutional networks and convolutional neural networks (BiTCN-CNN), addressing the challenges of low recognition accuracy and overfitting in small-sample training. The CatBoost algorithm selects key features from inertial measurement unit (IMU) data to perform locomotion mode recognition. EMD decomposes the signal into intrin-sic mode functions (IMFs) and progressively extracts the residuals, thereby enhancing the angular data while increasing the diversity and representativeness of the dataset. Subsequently, BiTCN-CNN models are employed to recognize various movement patterns. Motion data from 12 healthy subjects were collected, enhanced using EMD, and subsequently input into a BiTCN-CNN model for training and testing. The performance was compared with traditional recognition algorithms, such as Convolutional Neural Networks (CNNs), Temporal Convolutional Networks (TCN), and Long Short-Term Memory (LSTM) networks. The experimental results demonstrated that the method, combining EMD data augmentation and BiTCN-CNN, improved recognition accuracy under small sample conditions. Notably, for the five stable locomotion modes (level ground walking (LW), stair ascent (SA), stair descent (SD), ramp ascending (RA), and ramp descending (RD)), the recognition accuracy reached 99.27%. Furthermore, the model also demonstrated promising performance in eight pattern transition tasks (including LW-SA, LW-SD, LW-RA, LW-RD, SA-LW, SD-LW, RA-LW, RD-LW), achieving an accuracy of 97.27%.

Abstract Examining Lidar data is an efficient way to detect ancient Maya features across the Yucatan Peninsula. Automated object detection powered by deep learning leverages Maya archaeologists’ specialist knowledge in detecting the presence of ancient Maya settlements. By using a broadscale approach in its training, our new efficient multi-regional model Q2000 achieves comparable performance across a significantly broader and more diverse geographic region. This study addresses the current limitation of small-scale, area-specific models to generalize characteristics and properly detect a diverse range of target objects over a large area. This study introduces the foundational development of a broadscale, multi-region convolutional neural network (CNN) object detection model utilizing Lidar data across a significantly larger extent of the Maya area (approximately 35,584 km 2 ). This model achieved accuracies comparable to previous local studies that relied on the annotation of a larger number of structures within smaller, more homogeneous areas. Comparative analysis of the model's test results indicates enhanced generalization across diverse topographic regions when trained on multi-area data, achieving a robust F1 Score of 0.89, even with a relatively small training sample set. Our results further indicate that a broadscale approach to deep learning is efficient, and that a pan-Yucatan model can be effective.

Traditional optical systems often require multiple lenses and substantial physical space to achieve high imaging quality. In contrast, single-plane diffractive optical elements (SPDOEs) enable compact, single-element imaging due to their unique imaging characteristics. This paper proposes a prior-guided computational imaging method tailored for SPDOEs. Specifically, a prior-guided attention-enhanced multiscale deblurring network (PAMDN) was developed to achieve high-quality image reconstruction. An SPDOE with an easily manufacturable f-number of 5 and a focal length of 50 mm was designed and fabricated. A full field-of-view (FOV) average wavefront aberration model was established, and diffraction efficiency under fabrication errors was analyzed. Based on these, RGB-channel prior PSF features and loss-function weights were constructed. Frequency-domain features were extracted via inverse filtering of the degraded image and prior PSF. A multiscale attention mechanism then fuses both spatial- and frequency-domain features for final image reconstruction. A small-sample training dataset was established using real-world images captured by the fabricated SPDOE and a high-quality industrial lens. Experimental results show that PAMDN improves the average peak signal-to-noise ratio (PSNR) by 30% over conventional PSF-based deconvolution methods, with gains exceeding 13 dB across various FOV regions. This paper provides a theoretical foundation for high-quality imaging with SPDOEs and offers a new perspective for the miniaturization of future optical systems.

While deep convolutional neural networks (DCNNs) have achieved remarkable performance in chest X-ray interpretation, their success typically depends on access to large-scale, expertly annotated datasets. However, collecting such data in real-world clinical settings can be difficult because of limited labeling resources, privacy concerns, and patient variability. In this study, we applied a multimodal Transformer pretrained on free-text reports and their paired CXRs to evaluate the effectiveness of this method in settings with limited labeled data. Our dataset consisted of more than 1million CXRs, each accompanied by reports from board-certified radiologists and 31 structured labels. The results indicated that a linear model trained on embeddings from the pretrained model achieved AUCs of 0.907 and 0.903 on internal and external test sets, respectively, using only 128 cases and 384 controls; the results were comparable those of DenseNet trained on the entire dataset, whose AUCs were 0.908 and 0.903, respectively. Additionally, we demonstrated similar results by extending the application of this approach to a subset annotated with structured echocardiographic reports. Furthermore, this multimodal model exhibited excellent small sample learning capabilities when tested on external validation sets such as CheXpert and ChestX-ray14. This research significantly reduces the sample size necessary for future artificial intelligence advancements in CXR interpretation.

Using Artificial Intelligence for Text Screening in a Systematic Review of Cardiotoxicity

To address the small-sample training bottleneck and inadequate convergence efficiency of Deep Reinforcement Learning (DRL)-based communication anti-jamming methods in complex electromagnetic environments, this paper proposes a Generative Adversarial Network-enhanced Deep Q-Network (GA-DQN) anti-jamming method. The method constructs a Generative Adversarial Network (GAN) to learn the time–frequency distribution characteristics of short-period jamming and to generate high-fidelity mixed samples. Furthermore, it screens qualified samples using the Pearson correlation coefficient to form a sample set, which is input into the DQN network model for pre-training to expand the experience replay buffer, effectively improving the convergence speed and decision accuracy of DQN. Our simulation results show that under periodic jamming, compared with the DQN algorithm, this algorithm significantly reduces the number of interference occurrences in the early communication stage and improves the convergence speed, to a certain extent. Under dynamic jamming and intelligent jamming, the algorithm significantly outperforms the DQN, Proximal Policy Optimization (PPO), and Q-learning (QL) algorithms.

Abstract The study of advanced rotating machinery fault diagnosis technology is of great significance to improve the safety and reliability of equipment operation. Training samples are difficult to obtain in engineering practice, leading to weak generalization and low diagnostic accuracy of deep learning-based fault diagnosis methods. To solve the problem of insufficient small sample data to support the training of traditional intelligent diagnostic methods, a small sample fault diagnosis method based on frequency-guided multi-scale convolutional neural network (CNN) model is proposed in this paper. The model consists of multi-scale feature extraction (MFE), dual convolution fusion (DCF), frequency guided feature fusion (FGFF) module, and kolmogorov–arnold networks fully connected (KAN-FC) module. The MFE can comprehensively extract original fault signal features even with small sample data; the DCF effectively integrates the attention weight relationships between channels and spatial dimensions; and the FGFF integrates features extracted from multiple branches to build the feature network relationship. Additionally, the KAN is introduced as the fully connected layer of the model, which can better adapt to the differences in small data samples under various working conditions. When the proportion of the small sample training set is 0.3, the G-mean value on the Case Western Reserve University dataset is 99.89%, and the G-mean value on the Paderborn University (PU) dataset is 100%, which is approximately 0.59%–9.69% higher than that of other models. Through comparative verification, it is demonstrated that the proposed model outperforms existing models in small sample fault diagnosis and has strong generalization performance.

Detecting and segmenting damaged wires in substations is challenging due to varying lighting conditions and limited annotated data, which degrade model accuracy and robustness. In this paper, a novel 24 h × 7 days broken wire detection and segmentation framework based on dynamic multi-window attention and meta-transfer learning is proposed, comprising a low-light image enhancement module, an improved detection and segmentation network with dynamic multi-scale window attention (DMWA) based on YOLOv11n, and a multi-stage meta-transfer learning strategy to support small-sample training while mitigating negative transfer. An RGB dataset of 3760 images is constructed, and performance is evaluated under six lighting conditions ranging from 10 to 200,000 lux. Experimental results demonstrate that the proposed framework markedly improves detection and segmentation performance, as well as robustness across varying lighting conditions.

The precise prediction of transcription factor binding sites (TFBSs) is crucial in understanding gene regulation. In this study, deepTFBS, a comprehensive deep learning (DL) framework that builds a robust DNA language model of TF binding grammar for accurately predicting TFBSs within and across plant species is presented. Taking advantages of multi-task DL and transfer learning, deepTFBS is capable of leveraging the knowledge learned from large-scale TF binding profiles to enhance the prediction of TFBSs under small-sample training and cross-species prediction tasks. When tested using available information on 359 Arabidopsis TFs, deepTFBS outperformed previously described prediction strategies, including position weight matrix, deepSEA and DanQ, with a 244.49%, 49.15%, and 23.32% improvement of the area under the precision-recall curve (PRAUC), respectively. Further cross-species prediction of TFBS in wheat showed that deepTFBS yielded a significant PRAUC improvement of 30.6% over these three baseline models. deepTFBS can also utilize information from gene conservation and binding motifs, enabling efficient TFBS prediction in species where experimental data availability is limited. A case study, focusing on the WUSCHEL (WUS) transcription factor, illustrated the potential use of deepTFBS in cross-species applications, in our example between Arabidopsis and wheat. deepTFBS is publically available at https://github.com/cma2015/deepTFBS.

To tackle the challenges of relative attitude adaptability and limited sample availability in underwater moving target recognition for active sonar, this study focuses on key aspects such as feature extraction, network model design, and information fusion. A pseudo-three-dimensional spatial feature extraction method is proposed by integrating generalized MUSIC with range–dimension information. The pseudo-WVD time–frequency feature is enhanced through the incorporation of prior knowledge. Additionally, the Doppler frequency shift distribution feature for underwater moving targets is derived and extracted. A multidimensional feature information fusion network model based on meta-learning is developed. Meta-knowledge is extracted separately from spatial, time–frequency, and Doppler feature spectra, to improve the generalization capability of single-feature task networks during small-sample training. Multidimensional feature information fusion is achieved via a feature fusion classifier. Finally, a sample library is constructed using simulation-enhanced data and experimental data for network training and testing. The results demonstrate that, in the few-sample scenario, the proposed method leverages the complementary nature of multidimensional features, effectively addressing the challenge of limited adaptability to relative horizontal orientation angles in target recognition, and achieving a recognition accuracy of up to 97.1%.

With the intensification of global climate change, the frequent occurrence of typhoon disaster events has become a great challenge to the sustainable development of cities around the world; thus, it is of great significance to carry out the assessment of typhoon-directed economic losses. Typhoon disaster loss assessment faces key challenges, including complex regional environments, scarce historical data, difficulties in multi-source heterogeneous data fusion, and challenges in quantifying assessment uncertainties. Meanwhile, existing studies often overlook the complex relationship between the spatial expansion of urban and rural construction (SEURC) and typhoon disaster losses, particularly their differential manifestations across different regions and disaster intensities. To address these issues, this study proposes CLPFT (Comprehensive Uncertainty Assessment Framework for Typhoon), an innovative assessment framework integrating prototype learning and uncertainty quantification through a UProtoMLP neural network. Results demonstrate three key findings: (1) By introducing prototype learning, a meta-learning approach, to guide model updates, we achieved precise assessments with small training samples, attaining an MAE of 1.02, representing 58.5–76.1% error reduction compared to conventional machine learning algorithms. This reveals that implicitly classifying typhoon disaster loss types through prototype learning can significantly improve assessment accuracy in data-scarce scenarios. (2) By designing a dual-path uncertainty quantification mechanism, we realized high-reliability risk assessment, with 95.45% of actual loss values falling within predicted confidence intervals (theoretical expectation: 95%). This demonstrates that the dual-path uncertainty quantification mechanism can provide statistically credible risk boundaries for disaster prevention decisions, significantly enhancing the practical utility of assessment results. (3) Further investigation through controlling dynamic assessment factors revealed significant regional heterogeneity in the relationship between SEURC and directed economic losses. Furthermore, the study found that when typhoon intensity reaches a critical value, the relationship shifts from negative to positive correlation. This indicates that typhoon disaster loss assessment should consider the interaction between urban resilience and typhoon intensity, providing important implications for disaster prevention and mitigation decisions. This paper provides a more comprehensive and accurate assessment method for evaluating typhoon disaster-directed economic losses and offers a scientific reference for determining the influencing factors of typhoon-directed economic loss assessments.

This paper proposes a multi-frame dual-stream 2DCNN-LSTM model (MF-DS-2DCNN-LSTM) for automatic modulation recognition. The model discretizes long sequences into two-dimensional frame structures and uses 2D CNN and LSTM together to model the spatiotemporal features of multi-channel IQ/AP signals. By employing a frame-based strategy, the original signal is reshaped into small images, with the 2D CNN extracting intra-frame spatial structures and inter-channel interaction features, while the LSTM captures the temporal evolution between frames. This approach integrates hierarchical modeling concepts from image processing and video analysis, and utilizes the Crested Porcupine Optimizer for hyperparameter tuning. Simulations show that, when recognizing nine modulation types, the model significantly outperforms methods such as CLDNN, achieving an average accuracy of 91.4% under high-SNR conditions (SNR above 2 dB). Moreover, the model maintains an accuracy of over 90% in small-sample training scenarios for SNRs above 4 dB. After optimization with the Crested Porcupine Optimizer, the models performance improved by 2.2%, and a 20.7% reduction in parameters was achieved.

Context. Modern technical objects (in particular vehicles) are extremely complex and place high demands on reliability. This requires automation of condition monitoring and fault diagnosis of objects and their components. The predictive maintenance improves operational readiness of technical objects. The object of study is a technical object health and usage monitoring process. The subject of study is a methods of computational intelligence for data-driven model building and related data processing tasks for health and usage monitoring system.Objective. The purpose of the work is to formulate data processing problems, to form a data set for data-driven model building and construct simple method for automatic diagnostic model building on example of helicopter health and usage monitoring system.Method. The method is proposed for the mapping of multidimensional data into a two-dimensional space preserving local properties of class separation, allowing for the visualization of multidimensional data and the production of simple diagnostic models for the automatic classification of diagnostic objects. The proposed method allows obtaining highly accurate diagnostic model with small training samples, provided that the frequency of classes in the samples is preserved. A method for synthesizing diagnostic models based on a two-layer feed-forward neural network is also proposed, which allows obtaining models in a non-iterative mode.Results. A sample of observations of the state of helicopter gears was obtained, which can be used to compare data-driven diagnostic methods and data processing methods that solve the problems of data dimensionality reduction. The Software has been developed that allows displaying a sample from a multidimensional to a two-dimensional space, which makes it possible to visualize data and reduces the dimensionality of the data. Diagnostic models have been obtained that allow automating the decision-making process on whether the diagnosed object (helicopter gear) belongs to one of two classes of states.Conclusions. The results of conducted experiments allow to conclude that the proposed method provides a significant reduction in the data dimensionality (in particular, for the considered problem of constructing a model for helicopter gear diagnosis, it reducesthe data dimensionality due to the compression of features by 46876 times). As the results of the conducted experiments for randomly selected instances in a two-dimensional system of artificial features obtained on the basis of the proposed method showed a significant reduction of the sample for individual tasks may allow to provide acceptable accuracy. And taking into account individual estimates of the instance significance will allow, even for small samples, to ensure the topological representativeness of the formed sample in relation to the original sample. The prospects for further research are to compare methods for constructing data-driven models, as well as methods for reducing the dimensionality of data based on the proposed sample. Additionally, it may be of interest to study a possible combination of theproposed method with methods for sample forming using metrics of the value of instances.

In addressing the problem of indoor air pollution source localization, traditional methods have limitations such as strong sample dependence and low computational efficiency. This study uses a convolutional neural network to establish a pollution source inversion method based on small samples. By integrating computational fluid dynamics simulation data and deep learning techniques, a spatial pollution source identification model suitable for limited-sample conditions was constructed. In a benchmark scenario, the optimized model achieved a localization of 82.3% weighted accuracy within a prediction radius of 1 m, and the corresponding normalized error of the detected area was of less than 0.26%. In cross-scenario verification, the localization accuracy within a 1 m radius increased to 100%, and the corresponding predicted Euclidean distance error decreased by 21.43%. By using the optimal cutting ratio (α = 0.25) and a rotation-enhanced dataset (θ = 10°, n = 36), the model reduced the cross-space sample requirement to 1/5 of that of the benchmark scenario while ensuring the accuracy of spatial representation. The research findings provide an efficient and reliable deep learning solution for the localization of pollution sources in complex spaces.

The recently developed semi-parametric generalized linear model (SPGLM) offers more flexibility as compared to the classical GLM by including the baseline or reference distribution of the response as an additional parameter in the model. However, some inference summaries are not easily generated under existing maximum-likelihood-based inference (GLDRM). This includes uncertainty in estimation for model-derived functionals such as exceedance probabilities. The latter are critical in a clinical diagnostic or decision-making setting. In this article, by placing a Dirichlet prior on the baseline distribution, we propose a Bayesian model-based approach for inference to address these important gaps. We establish consistency and asymptotic normality results for the implied canonical parameter. Simulation studies and an illustration with data from an aging research study confirm that the proposed method performs comparably or better in comparison with GLDRM. The proposed Bayesian framework is most attractive for inference with small sample training data or in sparse-data scenarios.