Training Data Research Articles

Introduction to Generative Artificial Intelligence: Contextualizing the Future.

Generative artificial intelligence (GAI) is a promising new technology with the potential to transform communication and workflows in health care and pathology. Although new technologies offer advantages, they also come with risks that users, particularly early adopters, must recognize. Given the fast pace of GAI developments, pathologists may find it challenging to stay current with the terminology, technical underpinnings, and latest advancements. Building this knowledge base will enable pathologists to grasp the potential risks and impacts that GAI may have on the future practice of pathology. To present key elements of GAI development, evaluation, and implementation in a way that is accessible to pathologists and relevant to laboratory applications. Information was gathered from recent studies and reviews from PubMed and arXiv. GAI offers many potential benefits for practicing pathologists. However, the use of GAI in clinical practice requires rigorous oversight and continuous refinement to fully realize its potential and mitigate inherent risks. The performance of GAI is highly dependent on the quality and diversity of the training and fine-tuning data, which can also propagate biases if not carefully managed. Ethical concerns, particularly regarding patient privacy and autonomy, must be addressed to ensure responsible use. By harnessing these emergent technologies, pathologists will be well placed to continue forward as leaders in diagnostic medicine.

Synthetic Image Generation for Military Vehicle Detection

This research paper investigates the effectiveness of using computer graphics(CG) based synthetic data for deep learning in military vehicle detection. In particular, we explore the use of synthetic image generation techniques to train deep neural networks for object detection tasks. Our approach involves the generation of a large dataset of synthetic images of military vehicles, which is then used to train a deep learning model. The resulting model is then evaluated on real-world images to measure its effectiveness. Our experimental results show that synthetic training data alone can achieve effective results in object detection. Our findings demonstrate the potential of CG-based synthetic data for deep learning and suggest its value as a tool for training models in a variety of applications, including military vehicle detection.

Generative AI model privacy: a survey

The rapid progress of generative AI models has yielded substantial breakthroughs in AI, facilitating the generation of realistic synthetic data across various modalities. However, these advancements also introduce significant privacy risks, as the models may inadvertently expose sensitive information from their training data. Currently, there is no comprehensive survey work investigating privacy issues, e.g., attacking and defending privacy in generative AI models. We strive to identify existing attack techniques and mitigation strategies and to offer a summary of the current research landscape. Our survey encompasses a wide array of generative AI models, including language models, Generative Adversarial Networks, diffusion models, and their multi-modal counterparts. It indicates the critical need for continued research and development in privacy-preserving techniques for generative AI models. Furthermore, we offer insights into the challenges and discuss the open problems in the intersection of privacy and generative AI models.

Cone Beam Computed Tomography Image-Quality Improvement Using "One-Shot" Super-resolution.

Cone beam computed tomography (CBCT) images are convenient representations for obtaining information about patients' internal organs, but their lower image quality than those of treatment planning CT images constitutes an important shortcoming. Several proposed CBCT image-quality improvement methods based on deep learning require large amounts of training data. Our newly developed model using a super-resolution method, "one-shot" super-resolution (OSSR) based on the "zero-shot" super-resolution method, requires only small amounts of training data to improve CBCT image quality using only the target CBCT image and the paired treatment planning CT image. For this study, pelvic CBCT images and treatment planning CT images of 30 prostate cancer patients were used. We calculated the root mean squared error (RMSE), peak signal-to-noise ratio (PSNR), and structural similarity (SSIM) to evaluate image-quality improvement and normalized mutual information (NMI) as a quantitative evaluation of positional accuracy. Our proposed method can improve CBCT image quality without requiring large amounts of training data. After applying our proposed method, the resulting RMSE, PSNR, SSIM, and NMI between the CBCT images and the treatment planning CT images were as much as 0.86, 1.05, 1.03, and 1.31 times better than those obtained without using our proposed method. By comparison, CycleGAN exhibited values of 0.91, 1.03, 1.02, and 1.16. The proposed method achieved performance equivalent to that of CycleGAN, which requires images from approximately 30 patients for training. Findings demonstrated improvement of CBCT image quality using only the target CBCT images and the paired treatment planning CT images.

An adaptive thresholding approach for open set fault diagnosis

Abstract Cross-domain fault diagnosis using deep learning plays a critical role in ensuring the reliability and safety of mechanical systems. However, real-world industrial scenarios often involve unknown fault classes, which introduce significant challenges beyond environmental differences between training and testing phases. These unknown fault classes, which do not appear in the training data, create a cross-domain open set fault diagnosis problem where the target domain includes both known and unknown fault types with distinct distribution characteristics. Traditional domain adaptation methods that align source and target domains often overlook the spatial distribution of each class in the feature space, leading to potential negative transfer and misclassification of unknown faults. To address these challenges, this paper proposes a k-nearest neighbors based adaptive thresholding (KNNAT) method, which dynamically adjusts classification thresholds based on the spatial distribution of each class in the feature space. This approach effectively isolates unknown faults, reducing their impact on domain adaptation and improving the reliability of the diagnostic process. Extensive experiments on the publicly available CWRU bearing and PHM09 datasets demonstrate that the proposed KNNAT method outperforms other state-of-the-art methods, achieving higher accuracy and robustness in identifying known faults while successfully isolating unknown faults. These results highlight the potential of using the KNNAT method to enhance the reliability of mechanical systems in cross-domain fault diagnosis applications.

Machine visual perception from sim-to-real transfer learning for autonomous docking maneuvers

AbstractThis paper presents a comprehensive approach to enhancing autonomous docking maneuvers through machine visual perception and sim-to-real transfer learning. By leveraging relative vectoring techniques, we aim to replicate the human ability to execute precise docking operations. Our study focuses on autonomous aerial refueling as a use case, demonstrating significant advancements in relative navigation and object detection. We introduce a novel method for aligning digital twins using fiducial targets and motion capture data, which facilitates accurate pose estimation from real-world imagery. Additionally, we develop cost-efficient annotation automation techniques for generating high-quality You Only Look Once training data. Experimental results indicate that our transfer learning methodologies enable accurate and reliable relative vectoring in real-world conditions, achieving error margins of less than 3 cm at contact (when vehicles are approximately 4 m from the camera) and maintaining performance at over 56 fps. The research findings underscore the potential of augmented reality and scene augmentation in improving model generalization and performance, bridging the gap between simulation and real-world applications. This work lays the groundwork for deploying autonomous docking systems in complex and dynamic environments, minimizing human intervention and enhancing operational efficiency.

Surrogate modeling of pantograph-catenary system interactions

The smooth interaction between the pantograph and the catenary is crucial for the operational safety of railway vehicles. Coupled dynamic models of the pantograph–catenary system (PCS) constructed based on physical principles are important tools for analyzing their interactions; however, these models rely on accurate system parameters (such as stiffness, damping, and mass). Under actual operating conditions, the system parameters of the PCS exhibit time-varying characteristics and are difficult to measure, making it challenging for dynamic models to accurately represent the system’s behavior. Data-driven intelligent algorithms, with powerful feature extraction and nonlinear fitting capabilities, provide new approaches for solving system response prediction and state identification problems of the PCS. However, excessive reliance on large amounts of data for training may lead to poor generalization ability and pose challenges to model robustness, such as sensitivity to input noise or outliers. To address these issues, this paper proposes a surrogate model for the interaction of the PCS by integrating physical information with deep neural networks. The model introduces a novel neural operator that combines Transformers and convolutions (Convs), capable of capturing complex mapping relationships among various parameters within the dynamic model of the PCS in the frequency domain. A residual network incorporating physical information is designed to simulate the intricate correlations among system parameters. Additionally, a dynamic weighting balance algorithm is proposed to adjust the losses of different physical equations dynamically, ensuring the balance of physical information during training. The proposed model effectively performs response prediction and state identification of the PCS. It demonstrates excellent performance on both simulation and real-world data, providing new insights and methodologies for studying PCS interactions.

Deepfake Detection Model Combining Texture Differences and Frequency Domain Information

In recent years, public security incidents caused by deepfake technology have occurred frequently around the world, which makes an efficient and accurate deepfake detection model crucial. The existing advanced methods use the manipulation features in the image to realize the binary classification of real and fake images by training complex neural network models. However, these models rely on a single manipulation feature, and the detection accuracy of these methods will be greatly reduced when the forgery technology or image quality of the training data set and the validation data set are different. Inspired by the existing work, we propose a two-stream collaborative learning framework that combines spatial texture differences and frequency information. The average difference convolution (ADC) is designed to extract the spatial texture difference information of the image, and the gray image frequency-aware decomposition (GFAD) is used to extract the artifact information of the image in the frequency domain. At the same time, the ViT idea is combined with cross attention mechanism for feature fusion to comprehensively mine forged features in forged images. Experimental results show that the proposed model has good detection effects on three benchmark data sets. In terms of cross-dataset evaluation, the AUC on Celeb-DF dataset reaches 82.86%, which is better than the existing advanced methods.

Using multiplexed functional data to reduce variant classification inequities in underrepresented populations

BackgroundMultiplexed Assays of Variant Effects (MAVEs) can test all possible single variants in a gene of interest. The resulting saturation-style functional data may help resolve variant classification disparities between populations, especially for Variants of Uncertain Significance (VUS).MethodsWe analyzed clinical significance classifications in 213,663 individuals of European-like genetic ancestry versus 206,975 individuals of non-European-like genetic ancestry from All of Us and the Genome Aggregation Database. Then, we incorporated clinically calibrated MAVE data into the Clinical Genome Resource’s Variant Curation Expert Panel rules to automate VUS reclassification for BRCA1, TP53, and PTEN.ResultsUsing two orthogonal statistical approaches, we show a higher prevalence (p ≤ 5.95e − 06) of VUS in individuals of non-European-like genetic ancestry across all medical specialties assessed in all three databases. Further, in the non-European-like genetic ancestry group, higher rates of Benign or Likely Benign and variants with no clinical designation (p ≤ 2.5e − 05) were found across many medical specialties, whereas Pathogenic or Likely Pathogenic assignments were increased in individuals of European-like genetic ancestry (p ≤ 2.5e − 05). Using MAVE data, we reclassified VUS in individuals of non-European-like genetic ancestry at a significantly higher rate in comparison to reclassified VUS from European-like genetic ancestry (p = 9.1e − 03) effectively compensating for the VUS disparity. Further, essential code analysis showed equitable impact of MAVE evidence codes but inequitable impact of allele frequency (p = 7.47e − 06) and computational predictor (p = 6.92e − 05) evidence codes for individuals of non-European-like genetic ancestry.ConclusionsGeneration of saturation-style MAVE data should be a priority to reduce VUS disparities and produce equitable training data for future computational predictors.

Highly accurate, efficient, and fabrication tolerance-aware nanostructure prediction for high-performance optoelectronic devices

Despite extensive efforts to predict optimal nanostructures for enhancing optical devices, a more accurate, efficient, and practical method for nanostructure optimisation is required. In particular, fabrication tolerance is a promising avenue for significantly improving manufacturing efficiency; however, research in this area is limited. In this study, we introduce a practical approach for enhancing the performance of optoelectronic devices using an artificial intelligence (AI)-based nanostructure optimisation strategy. We optimised a support vector regression (SVR) model to capture the complex and nonlinear relationships between the transmittance and nanograting structure variables with the goal of improving optoelectronic devices. Our versatile model accurately predicted the continuous transmittance data with high precision (R2 = 0.995) using only 216 training data points. It can also make predictions under untrained conditions, thereby enabling the creation of a transmittance nanostructure contour map (R2 = 0.949). This method facilitates the design of nanostructures tailored to specific optical properties and provides valuable insights into fabrication tolerance. Through experimental validation, we identified an optimal nanograting structure with the highest transmittance in the visible-light spectrum. When integrated into optoelectronic devices such as organic light-emitting diodes (OLEDs) and organic solar cells (OSCs), their performance is significantly improved by increasing the light transmittance. Specifically, devices using the fabricated nanograting film exhibited a 17% improvement in external quantum efficiency (EQE) for solution-processed organic light-emitting diodes (SP-OLEDs) and a 10.7% improvement in power-conversion efficiency (PCE) for OSCs.

Measurement-Based Neural Network Technique for Modeling the Low-Frequency Electric Field Radiated Behavior of Satellite Units

In this work, a machine learning technique employing a measurement-driven artificial neural network (ANN) architecture is proposed as a solution to the precise determination of the position and the moment of equivalent electric dipoles for unit characterization. These dipoles are used to match the generated electric field from various sources inside the spacecraft, during space exploration missions. Various methodologies for unit characterization have been proposed in the literature, the most common being the heuristic approaches, least squares variants, method of auxiliary sources, etc. Contrary to the previous time-consuming post-process methodologies, the proposed electric dipole neural network (EDMnet) can offer a real-time characterization of the measured unit (Device Under Test) after a proper training stage, especially as a fast pre-compliance method. The network uses the electric field vector, measured at 14 discrete locations, as input and reports the position and moment of the electric dipole that best matches the measured fields. In this work, various ANN architectures are tested and compared in order to select the optimal EDMnet parameters for accurate source identification. It is shown that the size of the artificial training data affects the performance of the network. The proposed EDMnet can provide accuracy in mm-scale, with respect to dipole positioning, greater than 99% in dipole moment prediction.

Burned Olive Trees Identification with a Deep Learning Approach in Unmanned Aerial Vehicle Images

Olive tree orchards are suffering from wildfires in many Mediterranean countries. Following a wildfire event, identifying damaged olive trees is crucial for developing effective management and restoration strategies, while rapid damage assessment can support potential compensation for producers. Moreover, the implementation of real-time health monitoring in olive groves allows producers to carry out targeted interventions, reducing production losses and preserving crop health. This research examines the use of deep learning methodologies in true-color images from Unmanned Aerial Vehicles (UAV) to detect damaged trees, including withering and desiccation of branches and leaf scorching. More specifically, the object detection and image classification computer vision techniques area applied and compared. In the object detection approach, the algorithm aims to localize and identify burned/dry and unburned/healthy olive trees, while in the image classification approach, the classifier categorizes an image showing a tree as burned/dry or unburned/healthy. Training data included true color UAV images of olive trees damaged by fire obtained by multiple cameras and multiple flight heights, resulting in various resolutions. For object detection, the Residual Neural Network was used as a backbone in an object detection approach with a Single-Shot Detector. In the image classification application, two approaches were evaluated. In the first approach, a new shallow network was developed, while in the second approach, transfer learning from pre-trained networks was applied. According to the results, the object detection approach managed to identify healthy trees with an average accuracy of 74%, while for trees with drying, the average accuracy was 69%. However, the optimal network identified olive trees (healthy or unhealthy) that the user did not detect during data collection. In the image classification approach, the application of convolutional neural networks achieved significantly better results with an F1-score above 0.94, either in the new network training approach or by applying transfer learning. In conclusion, the use of computer vision techniques in UAV images identified damaged olive trees, while the image classification approach performed significantly better than object detection.

Are more data always better? – Machine learning forecasting of algae based on long-term observations

Bloom-forming algae present a unique challenge to water managers as they can significantly impair provision of important ecosystem services and cause health risks to humans and animals. Consequently, effective short-term algae forecasts are important as they provide early warnings and enable implementation of mitigation strategies. In this context, machine learning (ML) emerges as a promising forecasting tool. However, the performance of ML models is heavily dependent on the availability of appropriate training data. Consequently, it is essential to determine the volume of data necessary to develop reliable ML forecasts. Understanding this will guide future monitoring strategies, optimize resource allocation, and set realistic expectations for management outcomes. In this study, we used 30 years of fortnightly measurements of 13 different parameters from a lake in the English Lake District (UK) to examine the impact of training data duration on the performance of ML models for forecasting chlorophyll-a two weeks in advance. Once training data availability exceeded four years, a Random Forest model was found to consistently outperform naive benchmarks (mean absolute percentage error 16.4 % lower than the best-performing benchmark). With more than 5 years of training data, model performance generally continued to improve, but with diminishing returns. Furthermore, it was found that equivalent and, in some cases, better performance could be achieved by only using a subset of the most important input features. Additionally, it was found that reducing the sampling frequency had negative impacts on performance, both due to the reduced number of training observations available, and increased forecast horizon. Our findings demonstrate that for lakes ecologically similar to the study site, a consistent and regular sampling programme focused on monitoring a limited number of key parameters can provide sufficient observations for generating short-term algae forecasts after approximately five years of data collection. Importantly, this result provides justification for the initiation of new monitoring programmes for sites where algal blooms are a concern, and suggests that there are likely many pre-existing monitoring datasets which would be suitable for training algae forecast models.

Expert System for Assessing Anxiety Levels in Toxic Relationships Using the Case-Based Reasoning Method Based On The Web

Experiences in toxic relationships often trigger significant emotional stress and impact mental health disorders. This study aims to develop a web-based expert system using the Case-Based Reasoning (CBR) method to evaluate anxiety levels caused by toxic relationships. The system is designed to provide more specific treatment by accurately analysing patterns of disorders resulting from toxic relationships. The system's development follows the waterfall model. System testing was conducted using the black-box testing method, demonstrating that the system performs as expected. The results of manual calculations were compared with the system outputs and showed consistency. Testing using 300 cases—80% as training data (240 cases) and 20% as testing data (60 cases)—achieved an accuracy of 91.67%. The recommendations provided include initial steps to manage anxiety. This indicates that the CBR method effectively distinguishes anxiety levels based on similar cases. These findings contribute to clinical psychology by providing a technological tool for quickly identifying anxiety levels. For practitioners, this system can serve as an initial reference before further treatment, while for users, it offers easy access to understanding their mental condition. This system is expected to be an innovative solution supporting accessible mental health care.

Parallel development of object recognition in newborn chicks and deep neural networks

How do newborns learn to see? We propose that visual systems are space-time fitters, meaning visual development can be understood as a blind fitting process (akin to evolution) in which visual systems gradually adapt to the spatiotemporal data distributions in the newborn’s environment. To test whether space-time fitting is a viable theory for learning how to see, we performed parallel controlled-rearing experiments on newborn chicks and deep neural networks (DNNs), including CNNs and transformers. First, we raised newborn chicks in impoverished environments containing a single object, then simulated those environments in a video game engine. Second, we recorded first-person images from agents moving through the virtual animal chambers and used those images to train DNNs. Third, we compared the viewpoint-invariant object recognition performance of the chicks and DNNs. When DNNs received the same visual diet (training data) as chicks, the models developed common object recognition skills as chicks. DNNs that used time as a teaching signal—space-time fitters—also showed common patterns of successes and failures across the test viewpoints as chicks. Thus, DNNs can learn object recognition in the same impoverished environments as newborn animals. We argue that space-time fitters can serve as formal scientific models of newborn visual systems, providing image-computable models for studying how newborns learn to see from raw visual experiences.

Complexity, interpretability and robustness of GP-based feature engineering in remote sensing

Feature engineering is a crucial step in machine learning that provides better data for the learning algorithms to induce robust models, and this effort should be adapted to the capabilities of each algorithm. For example, classifiers that do not perform data transformations (e.g., cluster-based) perform better when the different classes are separated, typically requiring preprocessed data. Other models (e.g., decision trees) can perform several splits in the feature space, easily obtaining perfect results in training data, but have a higher risk of overfitting with unprocessed data. We use the rbd-GP and M3GP genetic programming algorithms to induce new features based on the original features, to be used by shallow and deep decision tree and random forest models. M3GP is wrapped around a learning algorithm, using its performance as fitness. This way, the induced features are adapted to the classifier, allowing us to compare the complexity of the features induced for the different classifiers. We measure the complexity of the induced features using several structural and functional complexity metrics found in the literature, also proposing a new metric that measures the separability of classes in the feature space. Like other authors, we use complexity as an interpretability metric, selecting three models to discuss and validate based on their performance and size. We apply these methods to remote sensing classification problems and solve two tasks that are hard due to the high similarity between the land cover classes: detecting cocoa agroforest and forecasting forest degradation up to one year in the future.

RuGECToR: rule-based neural network model for russian language grammatical error correction

Grammatical Error Correction is one of the core Natural Language Processing tasks. At the moment, the open source state-of-the-art Sequence Tagger for English is GECToR model. For the Russian language, this problem does not have solutions with the same good results due to the lack of labeled datasets, therefore we decided to contribute to the aforementioned task. In this research, we described the process of creating a synthetic dataset and training a model on it. We adapt GECToR architecture for Russian language and call it RuGECToR. We use this architecture because, unlike Sequence-to-Sequence approach, it is easy to interpret and does not require a lot of training data. Our goal was to train the model in such a way that it does not adapt to a specific sample, but generalizes the morphological properties of the language. The presented model achieves an F0.5 of 82.5 on synthetic data and an F0.5 of 22.2 on RULEC dataset which was not in the training set.

RESMatch: Referring expression segmentation in a semi-supervised manner

Referring Expression segmentation (RES), a task that involves localizing specific instance-level objects on the basis of free-form linguistic descriptions, has emerged as a crucial frontier in human–AI interactions. It demands an intricate understanding of both the visual and textual contexts, and often requires extensive training data. This paper introduces RESMatch, the first semi-supervised learning (SSL) approach for RES aimed at reducing reliance on exhaustive data annotation. Moreover, our proposed RESMatch can leverage the abundance of image–text paired data available in the current era of training large models, resulting in an improved RES task performance without the need for costly semantic annotation, as evidenced by our experimental results. Moreover, although existing SSL techniques are effective in image segmentation, we find that they fall short in RES. Facing challenges such as the comprehension of free-form linguistic descriptions and the variability in object attributes, RESMatch introduces a trifecta of adaptations: revised strong perturbation, text augmentation, and adjustments for pseudolabel quality and strong–weak supervision. RESMatch has demonstrated its effectiveness by achieving state-of-the-art (SOTA) results in various experimental settings. This pioneering work lays the groundwork for future research in semi-supervised learning for referring expression segmentation.

Advanced Charge Extended Hückel (CEH) Model and a Consistent Adaptive Minimal Basis Set for the Elements Z = 1-103.

The Charge Extended Hückel (CEH) model, initially introduced for adaptive atomic orbital (AO) basis set construction (J. Chem. Phys. 2023, 159, 164108), has been significantly revised to enhance accuracy and robustness, particularly in challenging electronic situations. This revision includes an extension toward f-elements, covering actinoids with their f-electrons in the valence space. We present a novel noniterative approximation for the electrostatic contribution to the effective Fock matrix, which substantially improves performance in polar or charged systems. Additionally, the training data set for elements Z = 1-103 has been expanded to encompass even more chemically diverse reference molecules as well as dipole moments and shell populations in addition to atomic charges. It includes a greater variety of "mindless" molecules (MLMs) as well as more complex electronic structures through open-shell and highly charged species. The revised method achieves mean absolute errors for atomic charges q of approximately 0.02 e- for randomly selected (mostly organic) molecules and 0.09 e- for MLMs, outperforming both classical charge models and established tight-binding methods. Furthermore, the revised CEH model has been validated through density functional theory calculations with the updated adaptive q-vSZP AO basis set on common thermochemical databases. Consistent with the extension of the CEH model, q-vSZP has also been variationally optimized and tested for elements Z = 58-71 and 87-103. The original versions of both CEH and q-vSZP are now considered deprecated.

Retinal Image Augmentation using Composed GANs

Medical image analysis faces a significant challenge in the scarcity of annotated data, which is crucial for developing generalizable Deep Learning (DL) models that require extensive training data. Consequently, the field of medical image generation has garnered substantial interest and potential for further exploration. Besides widely employed data augmentation techniques, such as rotation, reflection, and scaling, Generative Adversarial Networks (GANs) have demonstrated the ability to effectively leverage additional information from datasets by generating synthetic samples from real images. In the context of retinal image synthesis, an image-to-image translation approach is frequently adopted to generate retinal images from available vessel maps, which can be scarce and resource-intensive to obtain. Deviating from prior work reliant on pre-existing vessel maps, this study proposes a learning-based model that is independent of vessel maps, utilizing Progressive Growing GAN (PGGAN) to generate vascular networks from random noise. The visual and quantitative evaluations conducted suggest that the majority of the images generated by the proposed model are substantially distinct from the training set while maintaining a high proportion of true image quality, underscoring the model's potential as a powerful tool for data augmentation.