Search

Representation based classification (RC) algorithms have been successfully applied to pattern classification. However, most existing RC algorithms are not robust to bad training samples, since they ignore learning more discriminative dictionary atoms and instead directly use training samples as dictionary atoms. In addition, in order to improve expression ability and recognition accuracy, RC algorithms often need to expand the number of dictionary atoms, resulting in a surge in storage and computing costs. To obtain a more discriminative and compact dictionary, this study proposes discriminative dictionary learning for nonnegative representation based classification (DDLNRC). Specifically, in the DDLNRC, this paper utilizes nonnegative constraint to obtain a nonnegative representation for each training sample on the dictionary. In the dictionary learning stage, for a training sample, the DDLNRC minimizes the intra-class reconstruction error of the training sample and simultaneously enlarges the distance between the training sample and the atom that has the most influence on inter-class reconstruction error. Experiments demonstrate the effectiveness of the DDLNRC. Combined with the deep neural network features, it also can achieve higher accuracy than using the Softmax. • This paper proposes a DDLNRC for pattern recognition. • The DDLNRC can minimize intra-class reconstruction error. • The DDLNRC also can enlarge the inter-class reconstruction error. • By DDLNRC, a discriminative and compact dictionary can be obtained. • Experiments demonstrate the effectiveness of the DDLNRC on pattern recognition. [ABSTRACT FROM AUTHOR]

The emergence of artificial intelligence (AI) and its related capabilities has led industries to rethink the existing practices of conventional supply chain management and data analysis. Machine learning (ML), Deep Learning (DL) and their unique ability to predict future data and classify data have led to important research in the supply chain (SC) domain, particularly in identifying and prioritizing supply chain risks. This paper proposes several DL methodologies to exploit the benefit of DL, particularly to identify whether any product will be delivered late due to any unforeseen reason in a complex SC system. Four different DL architectures (Simple-LSTM, Deep-LSTM, 1D-CNN, and TCN-1DSPCNN models) are proposed to extract features, while six variant classifiers: Softmax, random trees (RT), random forest (RF), K-nearest neighbor (KNN), artificial neural network (ANN), and support vector machine (SVM), were used to classify delay or non-delay information. By seamlessly capturing intricate temporal dependencies, these DL models enhance accuracy in robustly identifying supply chain late orders. Leveraging their hierarchical feature learning, these proposed DL models excel in recognizing subtle patterns and correlations, making them ideal for classifying late orders within the supply chain. Their parallel processing prowess facilitates real-time decision support, allowing organizations to address potential delays and allocate resources effectively and proactively. Five-fold cross-validation is presented to avoid over-fitting and to prove the efficiency of the proposed DL models. The total accuracies of the six ML classifiers are 74.03, 75.81, 93.35, 87.72, 93.59, and 95.10, respectively, while the maximum accuracies obtained from four proposed DL methodologies obtained an accuracy of 97.6, 98.63, 100, 100% respectively using the SVM classifier for predicting late orders based on five-fold cross-validation. • This paper investigates a few DL approaches to extract features of SC data. • Both RNN and CNN are applied in the same model. • An improved CNN model has been proposed for feature extraction. • An online dataset is employed to validate the proposed DL architectures. [ABSTRACT FROM AUTHOR]

In the context of early childhood education, students need to acquire fundamental writing skills for their lifelong development. Public schools, especially in low- and middle-income countries, should monitor individual student progress to mitigate the detrimental effects of socioeconomic vulnerabilities in education. Furthermore, the volume of students often overwhelms teachers responsible for assessing handwriting texts and providing feedback. This article conducts a Systematic Literature Review (SLR) focusing on solutions for automatically evaluating students' handwriting, discussing their performance, future research directions, and areas needing further investigation. We used a mixed-methods approach to conduct our SLR, encompassing a search across four databases (ACM Digital Library, IEEE Xplore, ScienceDirect, and SpringerLink) and employed the snowballing technique. We used the initial set of papers from the database search as the foundation for the subsequent snowballing search. Findings revealed that the studies introduced computational techniques, examined or enhanced existing methods, and developed assessment tools. These solutions predominantly employed techniques such as artificial neural networks and used available datasets comprising handwritten images, answers, or student essays. Only some studies provide low-cost solutions for automatically assessing the writing abilities of underserved public school students. • We reviewed 491 papers, extracting data from 22 focusing on handwriting assessment. • We applied a mixed method using database search and the snowballing technique. • Studies frequently adopt deep learning methods to tackle the handwriting recognition problem. • Studies employ various methods to automatically assess text production, such as ANN, latent semantic analysis, content vector analysis, and CNN. • The mean accuracy for handwriting assessment was notably high at 93.64%, suggesting that the models in this group exhibited good performance. [ABSTRACT FROM AUTHOR]

Estimating the full state of physical systems, including thermal and flow status, from sparse measurements of limited sensors is a critical technology for perception and control. Neural networks have been used in recent studies to reconstruct the global field in a supervised learning paradigm. However, these studies encounter two major challenges: the lack of interpretability of black-box models and performance bottleneck caused by network structure and parameter optimization. This paper proposes a hybrid method based on proper orthogonal decomposition (POD) and deep neural networks (DNNs) to further enhance the interpretability and accuracy of flow and heat field reconstruction. The key idea is to leverage the inherent data modes extracted by POD that capture essential features in physical fields, and formulate the reconstruction problem as finding an optimal linear combination of dominant POD modes. To reduce the error introduced by underfitting and model structure, this paper estimates the coefficients of POD modes by establishing and solving a linear optimization problem that minimizes the gap between the recovered field and the exact measurements, rather than employing regression models. However, the underdetermined issue cased by the sparse measurements restricts the optimization problem to obtain a proper solution. To alleviate this problem, this paper presents to utilize the powerful non-linear approximation ability of DNNs to produce a reference field as auxiliary observations, which combines exact measurements to jointly constrain the optimization problem solving. Finally, the global physical field is reconstructed by superposing dominant POD modes weighted with the solved coefficients. By combining with POD technology, the proposed method can also improve the performance of neural networks on reconstruction problems with large-scale and irregular domains. The experiments conducted on the fluid and thermal benchmarks demonstrate that the proposed method can significantly boost neural network reconstruction performance and outperform existing POD-based methods. • A hybrid method based on POD and deep neural networks is developed for flow and heat field reconstruction from sparse measurements. • The data modes extracted by POD are explicitly used to alleviate black-box problems of neural networks. • The field reconstruction problem is converted to a linear optimization jointly constrained by exact measurements and DNN predictions. • The underdetermined issues with sparse measurements can be addressed. [ABSTRACT FROM AUTHOR]

As one of the most critical variables in the hot rolling process, the accuracy of rolling force prediction is directly associated with production stability and product quality. Purely data-driven approaches, however, are severely constrained by the quantity and quality of data, posing challenges for further enhancing the accuracy of rolling force prediction. In this paper, a theory fusion deep neural network (DNN) modelling approach was proposed and applied to the prediction of rolling force during hot plate rolling. In terms of model establishment, the novel NN structure was designed in consideration of the rolling mechanism, and senior variable inputs were added at shallow locations in the network to reduce the loss of critical information. In terms of model training, the method of using rolling theory to guide the initialization of the model was proposed to enable the model to learn the theoretical features more completely in the pre-training phase. Finally, a method to optimize the overall structure of the model using the sparrow search algorithm (SSA) was proposed to ensure the best prediction performance. The model was tested with the data in the developed platform, and the results indicated that the proposed method achieves the best accuracy and stability in this paper, and the response relationship between model inputs and output was consistent with existing theoretical knowledge. Thus, the model can be trusted and flexibly applied to the actual manufacturing processes. [ABSTRACT FROM AUTHOR]

Transformers are Deep Neural Networks (DNN) that utilize a self-attention mechanism to capture contextual relationships within sequential data. Unlike traditional neural networks and variants of Recurrent Neural Networks (RNNs), such as Long Short-Term Memory (LSTM), Transformer models excel at managing long dependencies among input sequence elements and facilitate parallel processing. Consequently, Transformer-based models have garnered significant attention from researchers in the field of artificial intelligence. This is due to their tremendous potential and impressive accomplishments, which extend beyond Natural Language Processing (NLP) tasks to encompass various domains, including Computer Vision (CV), audio and speech processing, healthcare, and the Internet of Things (IoT). Although several survey papers have been published, spotlighting the Transformer's contributions in specific fields, architectural disparities, or performance assessments, there remains a notable absence of a comprehensive survey paper that encompasses its major applications across diverse domains. Therefore, this paper addresses this gap by conducting an extensive survey of proposed Transformer models spanning from 2017 to 2022. Our survey encompasses the identification of the top five application domains for Transformer-based models, namely: NLP, CV, multi-modality, audio and speech processing, and signal processing. We analyze the influence of highly impactful Transformer-based models within these domains and subsequently categorize them according to their respective tasks, employing a novel taxonomy. Our primary objective is to illuminate the existing potential and future prospects of Transformers for researchers who are passionate about this area, thereby contributing to a more comprehensive understanding of this groundbreaking technology. • The paper presents a comprehensive survey on transformers for deep learning tasks. • The paper conducts a thorough analysis on highly effective models in five domains. • The paper classifies the models based on respective tasks using a proposed taxonomy. • The characteristics of the surveyed models are deeply explored and analyzed. • Future directions and challenges for transformer-based models are deciphered. [ABSTRACT FROM AUTHOR]

With the growing value and high training cost of deep neural network (DNN) models, there has been increasing research interest in deep neural network model watermarking. However, the existing literature lacks robust model watermarking methods with retrospective functionality, which hinders multi-user distribution and fine-grained copyright verification of deep neural network models. To achieve the model watermarking with retrospective functionality that is generally denoted the model fingerprinting, this study proposes a novel robust black-box fingerprinting scheme for deep classification neural networks, in which a fingerprint that is unique for each user is inserted in the given DNN. Specifically, a DWT-DCT-SVD-based embedding method is introduced to construct a high-quality poisoned image by spreading the fingerprint as a poisoned trigger to the entire image, which both leads to the traceability of the to-be-protected DNN and significantly enhances the stealthiness and anti-forgery of the embedded poisoned trigger. As the amplitude of the embedded fingerprint is rather small, it poses a challenge that makes the model hard to capture features of the embedded fingerprint during the training process. To tackle this challenge, this study then proposes a poisoned feature enhancement module, which groups classification categories of clean images without the fingerprint as one type while taking the poisoned images with the correct fingerprint and the poisoned adversarial images containing the other incorrect fingerprints as another two different types. This module, which can be removed during fingerprint verification, improves the reliability and fidelity of the model fingerprinting. Accordingly, an adversarial training strategy is designed, aiming at enhancing the effective training of classification neural network models that contain the fingerprint with small embedding strength. Combining these strategies then gives rise to the proposed scheme, which is denoted the UfNet for notational convenience. Extensive experimental simulation results demonstrate the excellent invisibility of the fingerprint in the poisoned image and the desirable verification of the fingerprint in case of even a one-bit difference. Additionally, it also exhibits robustness against the related attacks such as STRIP, Neural Cleanse, Fine-Pruning, and GradCAM, which outperforms the state-of-the-arts (SOTAs). Furthermore, the proposed UfNet is robust to collusion attacks and frame the innocent. These results show the feasibility and effectiveness of the introduced model fingerprinting framework and the proposed implementation methodology in this paper. • Introduce the concept and implementation of model fingerprinting. • Provide a digital-watermarking-based method for the poisoned image construction. • Develop a poisoned feature enhancement module and an adversarial training strategy. • Propose the UfNet for model fingerprinting, outperforming the state-of-the-arts. [ABSTRACT FROM AUTHOR]

One of the high-risk behaviors leading to severe traffic injuries is not wearing a seat belt. It is therefore very important to be able to automatically detect seat belts from surveillance images, encourage drivers to wear seat belts, and enhance passenger safety. In this paper, a novel deep neural network, Gated Bi-directional Long Short-Term Memory network with part-to-whole attention (GBL-PA), is proposed for seat belt detection from surveillance images. The innovation of our model lies in its unique diagonal sampling strategy, which meticulously captures the seat belt's fine details, typically oriented from top right to bottom left across the torso of vehicle occupants. Our framework's novelty is further encapsulated by the part-to-whole attention mechanism, which intelligently harmonizes the detailed local information from seat belt-specific patches with the broader contextual insights from the regional proposals. The pioneering design of a Gated Bi-directional LSTM network facilitates the dynamic integration of interactions across patches to deliver an optimized final prediction. The superiority of GBL-PA is established through rigorous comparison with the state-of-the-art methods on a new, large benchmark dataset comprising 14,936 images from traffic surveillance footage. Our framework demonstrates a notable improvement, achieving a mean Average Precision (mAP) of 72.3%, which surpasses the second best, YOLOX, by 0.9% mAP. This significant and consistent outperformance across various metrics underscores the transformative potential of GBL-PA in the realm of traffic safety enforcement. The source code of our framework is available at ANONYMISED. • Novel seat belt detection framework exploiting inherent properties of seat belts. • Patch Diagonal Sampling obtains patches to capture fine-grained seat belt details. • Part-to-whole Attention exploits attentive patch information under global context. • Gated Bi-LSTM captures interactions among patches to aggregate all information. • The framework greatly outperforms SoTA methods on the large, constructed dataset. [ABSTRACT FROM AUTHOR]

With rapid growth of online healthcare systems, chief complaint classification plays an important role in areas such as triage or doctor recommendation. Existing medical text classification techniques such as rule-based or learning-based methods fail to effectively utilize the inherent hierarchical structure of label descriptions that contain strong domain knowledge. In this paper, we propose a novel text classification framework for chief complaint by embedding both input text and hierarchical structure of label descriptions based on deep neural networks. The proposed framework makes use of not only three branches (i.e. chief complaint branch, main-category branch, and sub-category branch) with a Sequence Information Encoder to encode semantics from chief complaint and hierarchical structure of label descriptions but also a Hierarchical Relational Network with Attention module to capture complex relationships among them focusing on informative words with attentional scores. We evaluate our framework on two public medical datasets with label descriptions extracted from medical books and websites. Experimental results show that the proposed method outperforms other baseline techniques by a significant margin. The source code of our framework is available at ANONYMISED. [Display omitted] • Novel chief complaint classification model using hierarchical label descriptions. • SIE effectively embeds input text to capture sequential information with context. • HRNA captures relationships among chief complaint and category descriptions. • The framework greatly outperforms SoTA methods on two real world datasets. [ABSTRACT FROM AUTHOR]

Knowledge graph embedding has found widespread application across various fields due to its inherent structured data. However, some non-graph-based models necessitate improvement in accurately describing complex hierarchical structures and connecting intricate relational paths. Graph neural networks embedding methods provide a promising direction for exploring complex hierarchical modeling, facilitating the flow of information from nodes along relational paths. Nevertheless, the deficient interpretable characterization by the saliency estimation in entities, which contributes to the efficient performance of the model, is difficult to comprehend. This paper proposes a novel model to mitigate the described phenomenon by intensifying the original structure of knowledge graphs. Specifically, it integrates the Axiomatic Fuzzy Set entity semantic extraction framework with a heterogeneous relationship self-attention mechanism to bolster node representation. Next, this framework incorporates a convolutional neural network model to increase the interaction capabilities of entities and relations. This interpretable entity-level attention mechanism complements the structural information of the knowledge graphs, enabling more comprehensive neighbor information capture and improving the ability to identify the correct entity location. Experimental validation on two relationally complex datasets demonstrates superior performance compared to current methods, obtaining a new Hit@1 value of 45.6% on WN18RR. • Propose a hierarchical graph neural network knowledge representation framework. • Employ the Axiomatic Fuzzy Set to capture underlying semantics between entities. • The model has interpretable entity-level and heterogeneous relational attention. • Estimates of proximity values between nodes are yielded by semantic descriptions. • Produce promising performance compared to other knowledge graph embedding models. [ABSTRACT FROM AUTHOR]

Forecasting wind power is crucial to ensuring the electricity system runs steadily. However, wind power forecasting is hampered by the randomness and volatility of wind power data. In this paper, a deterministic and probabilistic decomposition ensemble model based on improved feature screening and optimal Gaussian mixed kernel function is proposed, which mainly includes feature screening, feature learning, intelligent weighted integration and interval prediction modules. First, an improved feature screening (IFS) is proposed to capture the optimal features of the original wind power data, eliminating the interference of redundant features. Then an optimized deep neural network is suggested to learn the outcomes of IFS and further explore the temporal correlation of these features. Then, considering the difference of the contribution of each feature component, the final result is obtained by integrating the preliminary forecast results using intelligent weighted integration (IWI). Finally, an optimal mixed kernel-based Gaussian process regression model is developed for interval prediction, which improves the adaptability of the model. The data sets of two wind farms in Inner Mongolia were used for empirical testing. On the two datasets, the model's root mean square error (RMSE) is 1.1845 and 1.7354, respectively, both of which are lower than the comparison model. The model has the lowest coverage width criterion (CWC) of 0.1499 and 0.1460 on the two datasets, respectively. [ABSTRACT FROM AUTHOR]

Nonlinear complementarity problems (NCP) have been extensively studied in optimization due to its widespread applications. In this paper, we utilize the neural dynamic approach to solve the NCP. By integrating the famous FB function an NR function, we construct a new type of complementarity functions with one parameter p , which is aesthetically pleasing and easy to apply. Combined with the Lagrange multiplier method, a new type of merit function is also developed. Based on the complementarity function and merit function, we transform the NCP into an unconstrained minimization problem. Then, by KKT condition and gradient descent method, we propose a Lagrange neural network method. Under mild conditions, every equilibrium point of the proposed neural network model is a solution of the NCP. More importantly, by throwing a delay factor, we also develop a novel delayed neural network model. Both of these networks are shown to be global convergent, Lyapunov stable and exponential stable. Finally, we give some numerical experiments of the two neural network approaches and also give some applications to the compressed sensing signal reconstruction. Simulation results indicate that the parameter p in the complementarity function plays an important role on the convergence rate of the two neural networks. The delayed neural network outperforms the non-delayed neural network in some specific situations. It also demonstrates that the two networks can efficiently reconstruct the original signals. • A new type of complementary functions is constructed. • A new neural network method is proposed. • We propose a delayed neural network and give its complete convergence analysis. • The two neural networks are applied to solve the CS efficiently. [ABSTRACT FROM AUTHOR]

• The LSP process is used to treat a Ti-6Al-4V thin representative specimen. • The FEM analysis is used to investigate the induced effects of LSP. • The induced effects are the residual stress, plastic strain and damage profiles. • The ANNs and the ANFIS approaches are used to predict the LSP induced effects. • The optimal conditions are determined and compared. This research article covers a comprehensive investigation on the prediction and optimization of residual stresses, plastic deformations and damage induced by laser shock peening (LSP) on a thin Ti-6Al-4 V leading edge sample of a turbine blade. The study is based on a two-part approach: the first part involves a numerical simulation of the LSP process and the characterization of its effects on the material. The finite element analysis (FEA) is used to simulate the complex interactions between laser parameters, material properties and mechanical response. This numerical analysis generates valuable insights into the residual stress distribution, the plastic deformation and the potential damage within the sample being treated. Based on the numerical results, the second part highlights the novel application of the artificial neural network (ANN) and the adaptive neuro-fuzzy inference system (ANFIS) methods. By taking advantage of the strength of ANN and ANFIS, the models are able to learn accurately from all the data generated by numerical simulations, allowing them to precisely predict and optimize the effects induced by LSP: residual stresses, plastic strains and damage. This paper provides a significant contribution to the field of materials science and engineering, by offering an in-depth comprehension of the LSP process impact on Ti-6Al-4 V leading-edge turbine blade specimens. The combined utilization of numerical analysis and artificial intelligence AI-driven techniques offers a comprehensive approach to optimize the LSP process, leading to the fatigue resistance improvement and service life extension of turbine blades and other critical aerospace components. The proven capability of ANN and ANFIS enables innovative solutions in aerospace engineering and surface treatment technologies, promoting the AI implementation in materials processing and design. [ABSTRACT FROM AUTHOR]

In the realm of decision-based attacks, which aim to mislead target models by manipulating output labels, the advent of the polar coordinates attack has marked a significant evolution. Distinguished by its utilization of sampled affine planes with unique geometrical properties, this method has demonstrated notable enhancements in query efficiency. However, existing implementations of the latest polar coordinates attack encounter two major problems: a contradiction in the update criteria and an overemphasis on allocating query resources to angle binary search. To address these issues, the Efficient Polar Coordinates Attack (EPCA) is introduced, a novel approach that resolves these concerns while significantly augmenting query efficiency. Furthermore, a mechanism termed Adaptive Activation Strategy is proposed to control the deployment of a recently presented plug-and-play module named Frequency Binary Search. The integration of these elements further enables EPCA to effectively circumvent the local optimum. Abundant experiments conducted on the ImageNet dataset demonstrate the exceptional improvements achieved by EPCA, establishing it as a state-of-the-art method in decision-based attacks. The simulation codes of this paper are accessible at https://github.com/RYC-98/EPCA. • Two problems in polar coordinates attacks are revealed. • Efficient Polar Coordinates Attack (EPCA) is used to solve these problems. • An Adaptive Activation Strategy with Frequency Binary Search is proposed. • EPCA attains the SOTA performance. [ABSTRACT FROM AUTHOR]

• A pioneering framework called Cross-to-Merge Training (C2MT) is proposed for learning with noisy labels (LNL). • We are the first to perform parameter aggregation on LNL after cross-training. • We introduce a novel class-balancing strategy, the Median Balance Strategy (MBS), which proves useful for sample selection. • C2MT demonstrates robustness with respect to hyper-parameters and network architectures. • Comprehensive experimental results substantiate the advantages of C2MT. The collection of large-scale datasets inevitably introduces noisy labels, leading to a substantial degradation in the performance of deep neural networks (DNNs). Although sample selection is a mainstream method in the field of learning with noisy labels, which aims to mitigate the impact of noisy labels during model training, the testing performance of these methods exhibits significant fluctuations across different noise rates and types. In this paper, we propose Cross-to-Merge Training (C2MT), a novel framework that is insensitive to the prior information in sample selection progress, enhancing model robustness. In practical implementation, using cross-divided training data, two different networks are cross-trained with the co-teaching strategy for several local rounds, subsequently merged into a unified model by performing federated averages on the parameters of two models periodically. Additionally, we introduce a new class balance strategy, named Median Balance Strategy (MBS), during the cross-dividing process, which evenly divides the training data into a labeled subset and an unlabeled subset based on the estimated loss distribution characteristics. Extensive experimental results on both synthetic and real-world datasets demonstrate the effectiveness of C2MT. The Code will be available at: https://github.com/LanXiaoPang613/C2MT. [ABSTRACT FROM AUTHOR]

This paper aims to propose a new hybrid Machine Learning (ML) with Egret Swarm Optimization Algorithm (ESOA) and Weighted Relevance-based Combination Strategy (WERCS) methods for predicting the patch loading resistance (PLR) of both longitudinally unstiffened and stiffened plate girders. A dataset of both stiffened and unstiffened plate girders under patch loading are collected and used for training and testing to generate the proposed models. Firstly, the WERCS method is used to resample the training set to improve the performance of training data. After that, four ML models, including Gradient Boosted Tree (GTB), Extreme Gradient Boosting algorithm (XGB), Artificial neural network (ANN), and CATBoost regression (CAT) are employed. The hyperparameter of all above ML models are optimized by using ESOA algorithm to choose the best performance models. It was found that the WERCS influences considerably the performance of hybrid ML models, and the WERCS + ESOA + XGB model presents the best performance compared to others. The accuracy of the WERCS + ESOA + XGB model is validated by comparing its predictive results with the existing design codes and formulae. Additionally, the Local Interpretable Model-Agnostic Explanations (LIME) method is used to assess the importance and contributions of each input variable in the proposed model. This algorithm can explain the global and local effects of features with each range and create favourable foundations for practical designing. A Graphical User Interface (GUI) tool is developed to conveniently estimate both the patch loading resistance of longitudinally unstiffened and stiffened plate girders. A Python library has been developed and uploaded to PyPi for more convenient use without the need for accessing the source code. [ABSTRACT FROM AUTHOR]

Learning the discriminative features of different faces is an important task in face recognition. By extracting face features with neural networks, it becomes easy to measure the similarity of different face images, which makes face recognition possible. To enhance a neural network's face feature separability, incorporating an angular margin during training is common practice. The state-of-the-art loss functions CosFace and ArcFace apply fixed margins between the weights of classes to enhance the inter-class separation of face features. Since the distribution of samples in the training set is uneven, the similarities between different identities are unequal. Therefore, using an inappropriately fixed angular margin may lead to problems such as that the model has difficulty converging or that the face features are not sufficiently discriminative. It is more intuitive to use adaptive angular margins are angular adaptive, which can increase as the angles between classes increase. In this paper, we propose a new angular margin loss named X2-Softmax. X2-Softmax loss has adaptive angular margins, which increase as the angle between different classes increases. The angular adaptive margin ensures model flexibility and effectively improves the effect of face recognition. We trained a neural network with X2-Softmax loss on the MS1Mv3 dataset and tested it on several evaluation benchmarks to demonstrate the effectiveness and superiority of our loss function. [ABSTRACT FROM AUTHOR]

This research paper presents a comprehensive experimental and statistical approach for the analysis of vibration amplitudes in a high-speed rotating shaft employing a squeeze film damper (SFD). The research combines a comprehensive analysis that connects input parameters and response parameters, with a special emphasis on vibration amplitudes along the X and Z axes. This research utilizes the rigorous procedures of Response Surface Methodology (RSM) together with a Box-Behnken design and harnesses the capabilities of Artificial Neural Network (ANN) optimization techniques. The variables under scrutiny encompass critical factors such as shaft rotational speed, extending up to 8000 rpm, oil pressure, with a range extending up to 100 bar, and various oil mix ratios, spanning from 10 % to 50 %. Various statistical measures are computed to assess the errors and coefficients of determination of the projected models. The artificial neural network (ANN) model has shown somewhat reduced prediction errors and a greater coefficient of determination compared to the Response Surface Methodology (RSM) for both the x and z axes of vibration amplitude. The values of mean absolute error (MAE), root mean squared error (RMSE), and coefficient of determination (R-squared) are found for both x and z axes from RSM (3.50, 3.77), (4.50, 4.72), and (0.81, 0.79), (1.88, 1.70), (2.41, 2.12), and (0.94, 0.95), respectively, using the ANN model. The overall mean absolute percentage error (MAPE) from the ANN model of both the x and z axes (9.73 %, 9.38 %) is found to be lower compared to the RSM model (18.18 %–20.50 %). The conclusive research reveals that the artificial neural network (ANN) prediction model outperforms the regression model based on Response Surface Methodology (RSM), exhibiting superior accuracy in predicting vibration amplitudes. The ANN approach is an excellent option for calculating vibration amplitudes in high-speed rotating shafts. Additionally, it provides significant benefits in terms of effectiveness and time economy. This research provides valuable new insights into the most efficient modeling approaches for vibration management and highlights the benefits of using Artificial Neural Networks (ANN) for predictive assessments in this context. Particle Swarm Optimization is used to minimize the vibration amplitude of the shaft along the x and z axes using experimental data. [ABSTRACT FROM AUTHOR]

• The paper explores AI decision-making to enhance transparency of CIN diagnosis using colposcopy images. • The study presents interpretable insights for clinical validation using different XAI methods. • Saliency method showed best result, using gradients to reveal likely malignancy traits. • Expert Clinic Validation of results conducted, along with statistical analyses. Although artificial intelligence techniques have performed well in analysing medical images, it is essential to consider the degree of understanding or interpretation in evaluations. This is especially important when diagnostic support is involved, as the trustworthiness of machine learning predictions needs to be clarified and evaluated. For deep learning algorithms to be translated into clinical practice, they must be made less opaque. To examine the classification of cervical dysplasia-diagnosed patients, we will explore explainable artificial intelligence (XAI) approaches, by implementing seven different interpretation techniques (Saliency, XRAI, Integrated Gradients, Smooth Gradients, Smooth Integrated Gradients, Grad-CAM, Smoothgrad-CAM). These XAI techniques shall be explored in order to assess the performance of four Convolutional Neural Network models (AlexNet, VGG16, MobileNet, and InceptionNet). Our results show that saliency qualities most closely match expert annotations, and there is a moderate to strong association between a model's sensitivity and the agreement between humans and computers. We also discovered a relevant relationship between expert insights and computational learning. This highlights the potential role of human expertise in influencing effective computer learning. [ABSTRACT FROM AUTHOR]

To enhance RUL prediction accuracy and uncertainty quantification, numerous methods have been developed, including model-based, data-driven, and hybrid approaches. However, model-based approaches struggle with complex relationships and uncertainties. Data-driven methods might overlook prior knowledge and struggle with limited data. Hybrid models for RUL prediction face two key challenges: inadequate use of physical information and difficulty in accurately quantifying uncertainty. Aiming at the problems of non-linearity, small sample sizes, and the absence of uncertainty quantification in RUL prediction, this paper introduces a hybrid method aimed at achieving RUL prediction and uncertainty quantification in few-slot scenarios. In this study, a hybrid approach that combines model-based approach and data-driven approach is proposed to achieve accurate RUL prediction. The uncertainty is measured based on a Bayesian framework. Specifically, the proposed method combines the degradation trend model using a double-exponential degradation model (DEDM), and the degradation fluctuations is predicted by a Gated Recurrent Unit (GRU) network to achieve point estimation of the RUL. Subsequently, an ensemble learning method is adopted to integrate the different modules using a Bayesian neural network (BNN) for uncertainty quantification. The applicability and effectiveness of the proposed method is investigated through case studies conducted on the three lithium battery datasets. The comparative analyses are also conducted with commonly used EMD-based approaches. Experimental results demonstrate that the proposed method has good performance in both data leakage and non-leakage scenarios and is more effective than individual methods to achieve accurate RUL prediction under small sample datasets. Finally, this study dedicates to integrate physical models and data-driven models to address the challenge of data drift in future research. [ABSTRACT FROM AUTHOR]

Recently, extensive research is showing that deep learning has enormous potential in image steganography. However, most steganography methods based on deep learning are not sufficiently invisible. In this paper, a novel end-to-end deep neural network for image steganography based on generative adversarial network (GAN) and discrete wavelet transformation (DWT) is proposed, which can greatly improve the invisibility of the method. Firstly, the cover image is transformed from the RGB domain to the DWT domain to embed information in frequency. Furthermore, a multi-scale attention convolution to fuse different scales of feature information is proposed, which can help the model to find a better location to embed the information. Finally, the fusion module is used to embed the information into the cover image at different modification intensities. In addition, a new Discriminator that compares cover images and stego images in a patch-to-patch way to improve the ability of the Discriminator. We analyze our algorithm from multiple aspects (e.g., security and invisibility) to verify the superior performance of the proposed method. To demonstrate generalization ability, experiments are conducted on three widely used datasets and show the superior performance of our method compared to other steganography methods based on deep learning. • End-to-end deep neural network based image steganography method. • An effective DWT based method to embed information in frequency. • A multi-scale attention convolution is used to find a better embedding location. • A fusion module is used to embed data at different modification intensities. • A new Discriminator that compares cover and stego in a patch-to-patch way. [ABSTRACT FROM AUTHOR]

A digital hologram (DH) is an ultra-high-value video content that includes three-dimensional information in two-dimensional data. Therefore, for the distribution of this content, the intellectual property rights must be protected. This paper proposes robust hologram watermarking based on deep neural networks using guided attack modules and attack rate control techniques. The proposed method robustly embeds the watermark into the original hologram to cope with various attacks while keeping the invisibility that the watermarked hologram is similar to the original hologram. Previous work considered PSNR during training. However, hologram exists in the complex domain. Thus SNR for complex numbers should be considered to confirm the reconstruction results. In addition, we propose an enhanced network with increased depth to improve performance compared with previous work. Experiments show that the proposed method outperforms the previous work regarding robustness and invisibility against various attacks, such as Crop Out, Salt & Pepper Noise, and Rotation. Therefore, this study is expected to be widely used in various fields, such as digital image authentication, copyright protection, and data security. [ABSTRACT FROM AUTHOR]

Influence maximization in social networks aims to identify users with the highest influence in the network and leverage them as initial spreaders to maximize the revenue of influence. Recently, an increasing number of studies have drawn attention to the study of influence maximization in multi-entity social networks, achieving some degrees of success. However, most of them still have two major shortcomings. First, most of the previous work oversimplifies the correlation between the target entity and other entities, and overlooks the influence of user's interest preference on information propagation. Second, the majority of existing methods face a trade-off between operational efficiency and solution accuracy, making it difficult to be effectively applied to large social networks. To fill this gap, this paper proposes a Deep neural network model based on Historical Behavior Sequences, named DHBS. DHBS consists of two core components: a Deep neural network model Incorporating Multi-head Self-attention, named DIMS; and a Reverse Influence Sampling algorithm based on Historical Behavior Sequence, named RIS-HBS. Specifically, DIMS captures the dynamic interest changes of users through its attention mechanism, enabling more accurate modeling of entity correlation in multi-entity social networks. RIS-HBS efficiently identifies a set of high-influential users for the target entity by incorporating the concept of reverse influence sampling. Experimental results on four publicly available network datasets demonstrate that DHBS outperforms baseline methods in terms of prediction accuracy, expected spread, and scalability, and DIMS plays a key role on RIS-HBS in the DHBS model. • Influence maximization is studied by mining users' historical behavior sequences. • Analyze entity correlation by capturing users' interest changes. • A neural attentive influence maximization model named DHBS is proposed. [ABSTRACT FROM AUTHOR]

A botnet is a group of hijacked devices that conduct various cyberattacks, which is one of the most dangerous threats on the internet. Individuals or organizations can effectively detect botnets by analyzing abnormal behaviors in network traffic. Existing works focus on extracting the deterministic behavioral features, which highly rely on statistical features and existing botnet interaction structures, resulting in unsatisfactory detection accuracy, especially for unknown botnet traffic. The botnet detection method based on the original traffic bytes has more advantages in this regard, especially the use of mining payload information in the traffic to enhance the identification of abnormal botnet behavior is the focus of this study. In this paper, we propose a dual-mode botnet detection scheme, which takes the original traffic bytes as the object, one is to encode the implicit semantic relationship between the traffic bytes through a multi-layer Transformer encoder, and the other is the network traffic Image representation, the spatial relationship of traffic bytes is captured by a deep neural network, and then botnet detection is achieved by maximizing the mutual information between the two. We conduct comprehensive experiments with both known botnets and unknown botnets to evaluate our scheme. Experimental results show that for known botnets, our approach achieves 99.84% and 91.92% detection accuracy with CTU-13 and ISCX-2014 datasets, respectively, which is 3.04% and 2.54% more accurate compared with the state-of-art (DL). For unknown datasets, our scheme is 10.19% more accurate than the existing traffic representation. • We extract the implicit semantic relationship of raw traffic for botnet detection. • We propose a dual-modal botnet detection method (Bot-DM) based on raw traffic. • We conduct extensive experiments to validate the superiority of the proposed methods. • We compare the detection performance of unknown botnets in different traffic modes. [ABSTRACT FROM AUTHOR]

Active learning is focused on minimizing the effort required to obtain labeled data by iteratively choosing fresh data samples for training a machine learning model. One of the primary challenges in active learning involves the selection of the most informative instances for labeling by an annotation oracle at each iteration. A viable approach is to develop an active learning strategy that aligns with the performance of a meta-learning model. This strategy evaluates the quality of previously selected instances and subsequently trains a machine learning model to predict the quality of instances to be labeled in the current iteration. This paper introduces a novel approach to learning for active learning, wherein instances are chosen for labeling based on their potential to induce the most substantial change in the current classifier. We explore various strategies for assessing the significance of an instance, taking into account variations in the learning gradient of the classification model. Our approach can be applied to any classifier that can be trained using gradient descent optimization. Here, we present a formulation that leverages a deep neural network model, which has not been extensively explored in existing learning-to-active-learn methodologies. Through experimental validation, our approach demonstrates promising results, especially in scenarios where there are limited initially labeled instances, particularly when the number of labeled instances per class is extremely limited. • A new meta-active learning strategy for deep neural network models. • The instance selection problem is modeled as a regression problem. • Four strategies of instance importance scoring by considering gradient variations. • Experiments have shown good results when few initially labeled objects are available. • A dimensionality reduction step for the instances improves the performances. [ABSTRACT FROM AUTHOR]

The objective of multi-focus image fusion (MFIF) is to generate a fully focused image through integrating multiple partially focused source images. Most of the existing methods do not fully consider the local gradient variation rate of the source image, which makes it difficult to accurately distinguish the small defocused (focused) region covered by the large focused (defocused) region. In addition, these methods also cause edge blurring because they do not take into account misregistration of the source images. To address these issues, in this paper, we propose a simple and effective multi-focus image fusion framework based on multi-channel Rybak neural network (MCRYNN) model. Specifically, the proposed MCRYNN model is highly sensitive to local gradient changes of images based on input receptive fields, which can process multiple source images in parallel and extract the features of focused regions. Moreover, the decision maps can accurately be generate in proposed method based on the information interaction effect of parallel network structure for multi-focus image fusion task. Finally, we conduct qualitative and quantitative experiments on public datasets, and the results show that the performance of the proposed method outperforms the state-of-the-art methods. • A primary attempt to use Rybak's visual cortex theory to perform the MFIF. • A novel multi-channel Rybak neural network (MCRYNN) is proposed. • A simple and efficient MFIF framework is proposed based on MCRYNN model. • The proposed method has a high computational efficiency. [ABSTRACT FROM AUTHOR]

Breast cancer is one of the serious diseases that threaten the life of many women worldwide. The seriousness of this disease is that it is often discovered in late stages after a period of its occurrence. This causes a wide spread of the disease and difficulty in its treatment. Another important characteristic of this disease is that it can return again after a period of its treatment. Therefore, predicting the occurrence or recurrence of such disease early is the best solution to have a high cure rate. The main objective of this paper is to improve the prediction performance of the breast cancer recurrence. Many methods have been proposed to predict breast cancer recurrence. However, these methods did not achieve the desired results on one of the most famous datasets in the field of breast cancer recurrence's prediction (i.e., Wisconsin Prognosis Breast Cancer (WPBC) dataset). The highest accuracy achieved using the previous methods is 89.89%. Therefore, this paper provides a framework for improving the prediction of breast cancer recurrence. The proposed framework has the ability to overcome many of the challenges in the existing dataset such as: (a) the problem of imbalance between classes using a data over-sampling technique; and (b) the large number of data dimensions using Principal Component Analysis (PCA), and a wrapper dimensionality reduction technique based on Genetic Algorithm (GA). It also uses the neural network algorithm to fuse the results of two individual classifiers (i.e., Random Forest (RF) and Support Vector Machine (SVM)). Our proposed framework evaluation showed a significant improvement in the predication performance. It achieved an accuracy of 98.3%, area under the curve of 99%, and precision, recall, and f1-measure of 98%. [ABSTRACT FROM AUTHOR]

Cardiovascular diseases are currently the major causes of death globally. Among the strategies to prevent cardiovascular issues, the automated classification of heart sound abnormalities is an efficient way to detect early signs of cardiac conditions leading to heart failure or other, even asymptomatic, complications, quite effective for timely interventions. Despite the significant improvements in this field, there are still limitations due to the lack of solutions, available data-sets and poor (mainly binary — normal vs abnormal) classification models and algorithms. This paper presents a Medical Cyber–Physical System (MCPS) for the automatic classification of heart valve diseases onsite, in a timely manner. The proposed MCPS, indeed, can be deployed into personal and mobile devices, addressing the limitations of existing solutions for patients, healthcare practitioners, and researchers, through an efficient and easy accessible tool. It combines different neural network models trained on a new Italian dataset of 132 adult patients covering 9 heart sound categories (1 normal and 8 abnormal), also validated against two main open-access (Physionet/CinC Challenge 2016 and Korean) datasets. The overall MCPS performance (time, processing and energy resource utilization) and the high accuracy of the models (up to 98%) demonstrated the feasibility of the proposed solution, even with few data. The dataset supporting the findings of this paper is available upon request to the authors. • DB of 132 heart sounds of Italian patients in 9 categories, available upon request. • MCPS for automated HVD classification via hierarchical ML model structures. • Compared six ML classifiers (RF, SVM, NB, LR, DNN, LSTM) for HVD and majority voting. • MCPS validated on two open access datasets (Physionet/CinC 2016 and Korean). • Time, processing and energy efficiency of MCPS validated, with usage guidelines. [ABSTRACT FROM AUTHOR]

Marine ecosystems and their fish habitats are becoming increasingly important due to their integral role in providing a valuable food source and conservation outcomes. Due to their remote and difficult to access nature, marine environments and fish habitats are often monitored using underwater cameras to record videos and images for understanding fish life and ecology, as well as for preserve the environment. There are currently many permanent underwater camera systems deployed at different places around the globe. In addition, there exists numerous studies that use temporary cameras to survey fish habitats. These cameras generate a massive volume of digital data, which cannot be efficiently analysed by current manual processing methods, which involve a human observer. Deep Learning (DL) is a cutting-edge Artificial Intelligence (AI) technology that has demonstrated unprecedented performance in analysing visual data. Despite its application to a myriad of domains, its use in underwater fish habitat monitoring remains under explored. In this paper, we provide a tutorial that covers the key concepts of DL, which help the reader grasp a high-level understanding of how DL works. The tutorial also explains a step-by-step procedure on how DL algorithms should be developed for challenging applications such as underwater fish monitoring. In addition, we provide a comprehensive survey of key deep learning techniques for fish habitat monitoring including classification, counting, localisation, and segmentation. Furthermore, we survey publicly available underwater fish datasets, and compare various DL techniques in the underwater fish monitoring domains. We also discuss some challenges and opportunities in the emerging field of deep learning for fish habitat processing. This paper is written to serve as a tutorial for marine scientists who would like to grasp a high-level understanding of DL, develop it for their applications by following our step-by-step tutorial, and see how it is evolving to facilitate their research efforts. At the same time, it is suitable for computer scientists who would like to survey state-of-the-art DL-based methodologies for fish habitat monitoring. • Marine environments and fish habitats are widely surveyed using underwater cameras. • Deep learning (DL) offers an effective solution for processing underwater videos. • We write a tutorial on the key concepts and steps of DL for underwater monitoring. • We provide a survey of key DL-based visual fish monitoring methods/datasets. • We shed light on challenges/opportunities of DL for underwater fish monitoring. [ABSTRACT FROM AUTHOR]

The conventional standardized theoretical models (such as Webster, Alcelik, Indo-HCM) used for the delay estimation revolve around the mathematical hypothesis and assumptions, making them static with limitations in accommodating the dynamic traffic behaviors. Recent advances in artificial intelligence make machine learning techniques suitable for estimating vehicle control delay compared to conventional methods. This paper demonstrates the application of several machine learning models developed by focusing on the fluctuations of traffic observed at an intersection having heterogeneous traffic conditions in Ahmedabad city for delay estimation. Several parameters are extracted from on-field and video surveys. However, not all parameters are relevant for accurately estimating vehicle control delay. Hence, a feature selection process consisting of several feature-scoring techniques from filter, wrapper, and embedded methods is applied to all the parameters. This process gave insights into the most statistically relevant independent parameters, and out of all the parameters, cycle time was found to be insignificant, with a feature score of 0 from almost all techniques. Hence, it was removed, and then the prominent parameters were then used to build a vehicle control delay model using support vector regression (SVR), K-nearest neighbor (KNN), artificial neural network (ANN), random forest regression (RF), and decision tree regression (DT). Error distribution, standard deviation of errors, coefficient of Determination (R2), and Root Mean Squared Error (RMSE) are the parameters used for evaluating the performance of the machine learning models. RF outperformed them all with a standard deviation of errors, R2, and RMSE of 11.065, 0.926, and 11.081 on testing data. But ANN, KNN, and DT also performed satisfactorily. Compared with conventional standardized theoretical models, all the machine learning models except SVR performed better. [ABSTRACT FROM AUTHOR]

Medical image analysis, particularly the segmentation of White Blood Cells (WBCs), holds critical importance in scrutinizing the quantity and morphology of these cells in smear images, a pivotal step in disease detection. Existing methodologies often fall short in differentiating between healthy and diseased WBCs, emphasizing color components over internal organizational nuances and external morphologies. This paper presents a pioneering contribution through the introduction of an encoder-decoder deep neural network with an indeterminacy fusion-based model explicitly designed for WBC extraction from blood images. In contrast to prior studies, our approach addresses the inadequacies by delving into the intricacies of WBC structure, characterizing object indeterminacy within the neutrosophic set domain. The encoder-decoder network is fortified by integrating WBC indeterminacy as a fusion component, enabling the segmentation of WBCs into distinct nucleus and cytoplasm regions. The efficacy of our methodology is evaluated across three datasets with varying resolutions. Quantitative metrics such as segmentation accuracy and intersection over union (IoU) are employed to assess the proposed network's performance, which consistently outperforms three original encoder-decoder networks. Notably, our model achieves commendable precision rates, with the highest mean segmentation accuracy reaching 0.95301 across the three datasets. These results underscore the enhanced performance of our segmentation network, attributing its success to the influential role of indeterminacy fusion in augmenting conventional segmentation methods. In summary, this work not only introduces a novel solution for WBC segmentation but establishes its superiority over existing methods. The interpretive insights gained from the results highlight the potential of indeterminacy fusion to drive advancements in the realm of medical image segmentation, emphasizing originality, significance, and superior performance metrics. [ABSTRACT FROM AUTHOR]

Research on automatic translation of sign language to verbal languages has been progressively explored in recent years to assist speech and hearing-impaired people in communicating with non-signers. In this paper, a tiny machine learning (TinyML) solution is proposed for sign language recognition using a low-cost, wearable, internet-of things (IoT) device. A lightweight deep neural network is deployed on the edge device to interpret isolated signs from the Indian sign language using the time-series data collected from the motion sensors of the device. The scarcity of labeled training data is addressed by employing the deep transfer learning approach. Here, the knowledge gained from the data collected using the motion sensors of a different device is used to initialize the model parameters. The performance of the model is assessed in terms of classification accuracy and prediction time for different sampling rates and transferring schemes. The model achieves an average accuracy of 87.18% when all the parameters are retrained with just 4 observations of each sign recorded from the motion sensors of the proposed IoT device. The recognized sign is transmitted to a cloud platform in real-time. A mobile application, SignTalk , is also developed, which wirelessly receives the predicted signs from the cloud and displays it as text. Additionally, text-to-speech conversion is also provided on SignTalk to vocalize the predicted sign for better communication. [ABSTRACT FROM AUTHOR]

Skin diseases are reported to contribute 1.79% of the global burden of disease. The accurate diagnosis of specific skin diseases is known to be a challenging task due, in part, to variations in skin tone, texture, body hair, etc. Classification of skin lesions using machine learning is a demanding task, due to the varying shapes, sizes, colors, and vague boundaries of some lesions. The use of deep learning for the classification of skin lesion images has been shown to help diagnose the disease at its early stages. Recent studies have demonstrated that these models perform well in skin detection tasks, with high accuracy and efficiency. Our paper proposes an end-to-end framework for skin lesion classification, and our contributions are two-fold. Firstly, two fundamentally different algorithms are proposed for segmenting and extracting features from images during image preprocessing. Secondly, we present a deep convolutional neural network model, S-MobileNet that aims to classify 7 different types of skin lesions. We used the HAM10000 dataset, which consists of 10000 dermatoscopic images from different populations and is publicly available through the International Skin Imaging Collaboration (ISIC) Archive. The image data was preprocessed to make it suitable for modeling. Exploratory data analysis (EDA) was performed to understand various attributes and their relationships within the dataset. A modified version of a Gaussian filtering algorithm and SFTA was applied for image segmentation and feature extraction. The processed dataset was then fed into the S-MobileNet model. This model was designed to be lightweight and was analyzed in three dimensions: using the Relu Activation function, the Mish activation function, and applying compression at intermediary layers. In addition, an alternative approach for compressing layers in the S-MobileNet architecture was applied to ensure a lightweight model that does not compromise on performance. The model was trained using several experiments and assessed using various performance measures, including, loss, accuracy, precision, and the F1-score. Our results demonstrate an improvement in model performance when applying a preprocessing technique. The Mish activation function was shown to outperform Relu. Further, the classification accuracy of the compressed S-MobileNet was shown to outperform S-MobileNet. To conclude, our findings have shown that our proposed deep learning-based S-MobileNet model is the optimal approach for classifying skin lesion images in the HAM10000 dataset. In the future, our approach could be adapted and applied to other datasets, and validated to develop a skin lesion framework that can be utilized in real-time. [ABSTRACT FROM AUTHOR]

Sequential recommendation aims to predict future user interactions by analyzing dynamic patterns within their historical behavior sequences. Deep neural networks have recently become popular for learning representations of these sequences in the time domain. However, representing users' intentions in the time domain faces challenges such as noise in interactions and sparsity of data. Contrastive learning and representation learning in the frequency domain can mitigate these issues from different perspectives. In this paper, to fully integrate time-domain sequence representations, frequency-domain sequence representations and contrastive learning based on them, we propose a model called T ime– F requency C onsistency based contrastive learning for S equential Rec ommendation (TFCSRec). TFCSRec utilizes a time-domain encoder with a fully connected network and a filter network to extract high-order features and catch pure sequential patterns. Then, a learnable frequency-domain encoder with a recurrent neural network is designed to capture sequential characteristics in the frequency-domain space. Finally, TFCSRec combines a recommendation task and two contrastive learning tasks to optimize the two user representation encoders. Its contrastive learning is designed to minimize a contrastive regularization loss and a time–frequency consistency loss, which for the first time is constructed directly on the time-domain sequence representation and the frequency-domain sequence representation. Experiments on five benchmark datasets show that the proposed TFCSRec model outperforms other sequential recommendation models based on deep neural networks. • Design a time-domain encoder with a filter to extract sequential patterns. • Devise a learnable frequency-domain encoder to capture sequential characteristics. • Propose a time–frequency consistency based contrastive loss for the first time. [ABSTRACT FROM AUTHOR]

• The proposed method had fused multi-level information compensation and U-Net. • The module designed for multi-level information compensation and compression. • The method had fused compressed features and correlation features. Recently, Deep Convolutional Neural Networks have demonstrated high-quality reconstruction in image super-resolution procedure. In this paper, we propose improved image super-resolution reconstruction via multi-level information compensation and U-Net network to address the problem that the image super-resolution reconstruction algorithm based on deep neural networks tends to lose feature information in the feature extraction process, resulting in the lack of texture and edge details in the reconstructed image. Firstly, we design the U-net like network for image super-resolution reconstruction, which performs multi-level feature extraction and channel compression for the input features through the down-channel branch. It fuses the compressed features and extracts the correlation features of different channels through the bottom module, and performs multi-level feature extraction and channel recovery for the compressed correlation features through the up-channel branch. The multi-level information compensation model is then designed to compensate for the information lost in the channel compression process and the information that is difficult to recover in the channel recovery process of U-net like network. The experimental results can show that the proposed algorithm achieves a significant improvement in Peak Signal-to-Noise Ratio and Structure Similarity Index and visual effect compared with the state-of-arts algorithms. The average experimental results of PSNR from proposed method had improved by 1.63 dB, 1.53 dB, 0.97 dB and 0.94 dB compared to SRCNN, HAT, DAT and CARN, respectively. [ABSTRACT FROM AUTHOR]

Remote sensing applications, such as detection and recognition of objects, need a large number of high spatial resolution (HR) images. As a feasible technique, image fusion is considered to produce HR images. However, most of remote sensing image fusion methods based on deep neural network (DNN) are mainly limited by the extraction of spatial and spectral features. These methods ignore the redundancy among these features, which causes spatial and spectral distortions in the fused image. In this paper, we propose a novel image fusion method based on a triple disentangled network (TDNet) with dual attention to reduce the redundancy in the extracted features. In the proposed method, it is assumed that the information in panchromatic (PAN) and low spatial resolution multispectral (LRMS) images can be encoded as the spatial, spectral, and common features. Specifically, we construct a triple-stream network to extract these features. To efficiently model the spatial and spectral information in PAN and LRMS images, local–global attention and interdependency attention are designed and integrated into the network. Then, the redundancy among these features is reduced by disentangled learning, in which these features are recombined to reconstruct the PAN and LRMS images. Besides, these features should complement each other. So, we utilize the maximal coding rate reduction to balance the redundancy and complementarity among them. Finally, all features are recombined to synthesize the high spatial resolution multispectral image. The experimental results demonstrate that the proposed TDNet has a superior performance in terms of qualitative and quantitative evaluations. The code link is https://github.com/RSMagneto/TDNet. • A triple disentangled network with dual attention is proposed for image fusion. • MCR loss is designed to disentangle spatial, spectral and common features in images. • The local–global spatial similarity is captured by the proposed SLGA module. • The spectral correlation among the LRMS image is modeled by the proposed SID module. [ABSTRACT FROM AUTHOR]

Fine-tuning and mask-tuning (or prompt tuning) are two approaches to construct deep neural networks for entity and relation extraction. Fine-tuning based models optimize neural networks with task-relevant objective, in which pre-trained language models (PLMs) are mainly used as external resources to support word embedding. In mask-tuning models, neural networks is optimized by the same pre-training objective in a PLM, which directly outputs verbalized entity type representations. It is effective to utilize potential knowledge of PLMs. In this paper, we propose a recollect-tuning approach, which jointly makes full use of the mechanisms of both fine-tuning and mask-tuning. In this approach, the recollect-tuning iteratively masks tokens in a possible entity span. The classification is based on both the masked token representation and the entity span representation. It is the same as the process to make a decision based on incomplete information. In the training process, the deep network is optimized by task-relevant objective, which strengthens the semantic representation of each entity span. It is effective to learn entity noise-invariant features and take full advantage of potential knowledge of PLMs. Our method is evaluated on three public benchmarks (the ACE 2004, ACE 2005 and SciERC datasets) for the entity and relation extraction task. The result shows significant improvement in the two tasks, outperforming the state-of-the-art performance on ACE04, ACE05 and SciERC by +0.4%, +0.6%, and +0.5%, respectively. • A unified framework combines fine-tuning and mask-tuning for NER and RE extraction. • A recollect-tuning approach proposed can learn noise-invariant entity features. • A grouped mask attention is designed to model the inter-span token interaction. [ABSTRACT FROM AUTHOR]

Optical Character Recognition (OCR) is a method to convert a scanned photo of"handwritten character recognition (HCR) or printed character recognition (PCR)" into a form of digital text. HCR is a version of OCR, which is remarkably modeled to identify handwritten text, while PCR aims at printed text identification. The identification of handwritten characters and digits is more complicated as compared to PCR because of the diversities in human writing styles, stokes, thickness, and curves of characters. Similarly, achievements in several computer vision tasks consider the Convolutional Neural Networks (CNN) to give an end-to-end solution for HCR with huge success. However, the process of significant feature learning for the identification of images is complicated with little data. Hence, this paper aims to develop a new handwritten character and digit recognition model with the incorporation of a deep learning strategy. Initially, the data related to Indian languages are collected from the standard benchmark datasets. Then, the collected data are given into the feature extraction phase 1, where the ResNet 151 is used for extracting the feature set 1. Similarly, the data gathered are considered in the feature extraction phase 2, where the Optimal Ensemble Pattern extraction approach is developed with Local Binary Pattern (LBP), Local Gradient Patterns (LGP), Local Tetra Pattern (LTrP), and Local Vector Pattern (LVP) for extracting the significant patterns from the language data. These extracted patterns are given into the ResNet 151 for getting the feature set 2. Here, the features from ResNet 151 get optimized with the enhanced optimization algorithm with Fitness-based Sail Fish Optimizer (F-SFO). The obtained feature set 1 and optimal feature set 2 are concatenated for performing final recognition. At last, the HCR is done with the help of developed Bi-LSTM-DNN to achieve the enhanced and accurate recognition of handwritten characters of the Indian languages. The performances of character recognition are further improved with the parameter optimization in Bi-LSTM-DNN with the same enhanced F-SFO. Overall result analysis, the accuracy of the designed method attains 95.12%. [ABSTRACT FROM AUTHOR]

• A scheme to recognize the distraction level was proposed. • MLP-AutoEncoder and DPC algorithms were used to define distraction-level labels. • Factors affecting driver distraction level were explored with mixed-model analyses. • A distraction-level prediction model with stacking ensemble learning was made. Driver distraction-level recognition while performing secondary tasks in full-touch in-vehicle information systems (FTIVISs) is essential for the harmonious co-driving of human and intelligent vehicle systems. However, there has been little research on this topic. To respond to this issue, this paper proposes a distraction-level recognition framework with a combination of semi-supervised learning, unsupervised learning, and supervised learning. First, unsupervised learning and semi-supervised is used to divide the collected unlabeled samples of driving distraction behavior into three categories of distraction levels: high, medium, and low. Second, the factors influencing the distraction level are explored through a mixed model analysis. Finally, a stacking-based ensemble learning model is proposed to recognize the driver distraction level by supervised learning, with the influencing factors of the distraction level used as model input parameters. To improve the recognition performance of the stacking model, four heterogeneous classifiers selected as the base classifiers, and principal component analysis (PCA) is introduced in the base classifier layer, which improves the ability of the meta -classifier to process raw features based on overfitting resistance. We conducted a real road experiment under different road and FTIVIS task conditions, and the proposed model performed better than traditional machine learning models. In addition, the model exhibited the greatest advantage when the deep neural network (DNN) algorithm was used as the meta -classifier of the model, with a recognition accuracy of 92.5%. The study findings are significant for developing a human–machine co-driving control strategy and improving vehicle driving safety. [ABSTRACT FROM AUTHOR]

Lithology identification is a crucial task for reservoir characterization and evaluation. There exists an intricate non-linear response between formation lithology and logging data. However, it is difficult to avoid lithology mislabeling due to human error and interpretation coarsening, and label quality can seriously affect the effectiveness of supervised learning. The presence of noisy labels makes it essential to learn with noisy labels. Noise-filtering methods and noise-robust algorithms only concentrate on a singular aspect of data or algorithm. In this paper, hybrid noise label filtering and correction framework for lithology identification (HNFCL) is proposed. Isolation forest is utilized to detect suspicious data, as it is efficient and fast. Baseline classifiers are built by ensemble tree models. In particular, the labels of abnormal data are removed and Tri-training semi-supervised method is introduced to relabel these data, which minimizes the loss of valid training data. Comprehensive experiments of the HNFCL framework, noise filtering methods and deep neural network methods with optimized loss functions were carried out in the industrial application of logging lithology identification. HNFCL achieved average accuracy of 87.94% and 94.93% in two study wells. These results outperformed the noise filtering methods and showed no significant difference from the state-of-the-art method. The correction of noise by HNFCL will provide a prospect for lithology identification applications. [ABSTRACT FROM AUTHOR]

Emergency Department overcrowding is a well-known problem. The consequences are long waiting times for patients, reduced service quality, and the potential for increased mortality rates. Moreover, this problem is escalating with the aging of the population and the chronification of diseases. Being aware of such concerns, Emergency Departments started to conduct patient accountability years ago, and now they are ready to take advantage of the gathered data to improve the quality of service to citizens. However, to achieve this goal, they need the technology that enables the generation of predictive models from such data to improve the planning of their resources. In this paper, we face the task of forecasting patient admissions by using Deep Neural Networks, particularly a modern Attention-based model. As well, historical records do not provide a full picture of how patients attend the emergency service. Relying on the possibilities given by the attention mechanism, we also propose the use of exogenous information such as calendar data, weather, air quality, allergens, and information extracted from the web via Google Trends to enhance admissions prediction accuracy. In this regard, we have tested different configurations of the exogenous variables to isolate which ones provide relevant information that improves the model. In our experiments, we have seen that the calendar, weather, and air quality provide the most valuable information, meanwhile, allergens and Google Trends data appear to be hindering the models' performance. • Using Attention model for forecasting Emergency Department (ED) patient admissions. • Comparing Attention model with LSTM-based networks for ED patients admissions. • Exogenous variables: weather, calendar, air quality, allergens and Google Trends. • Analyzing the impact of each type of exogenous variable in the forecast. • Attention model with calendar data achieves the best results. [ABSTRACT FROM AUTHOR]

The enhanced answer evaluation system is a cutting-edge automated tool that evaluates subjective answers in various contexts, such as educational assessments, surveys, and feedback forms. The proposed system leverages Natural Language Processing (NLP) and deep learning techniques to analyse subjective answers and provide evaluation scores with precision. Students' answers are evaluated based on various criteria, such as keywords, context, relevance, coherence, and similarity. This paper introduces an architecture for a subjective answer evaluator using three main aspects: detection of keywords, similarity matrix, and presence of named entities. It combines the three aspects and provides a final score. It provides a standardized mechanism to score a given user answer compared to the particular model answer without human prejudice. This research aims to transcend traditional methodologies that predominantly utilize keyword or keyphrase scoring (text-based similarity) to determine the final score of an answer without delving into its technical intricacies. The semantic similarity (vector-based) employs vector data representations for score calculation. This approach necessitates partitioning data into multiple vectors for a comprehensive analysis. While text similarity is effective for short answers, its efficacy diminishes as the length of the answer increases. Therefore, this study emphasizes the critical role of similarity scoring and Named Entity Recognition (NER) scoring in evaluating more extended responses based on the stsb-en-main dataset (short answers) and a custom dataset with 190 records. This research reveals its remarkable performance, which excels through a dynamic three-pronged approach: keyword scoring, semantic similarity, and NER scoring with models like Yet Another Keyword Extractor (YAKE), SimCSE and Camembert. These three independent components synergize to produce unmatched results, establishing a new standard in the field. This enhancement led to Root Mean Square Error (RMSE) scores of 0.031 (optimized error rate) and an impressive 71%+ accuracy for our comprehensive system. This achievement surpasses existing works, which typically reached accuracies ranging between 40%–60% for long answers. • Evaluation using a three-pronged approach of Keyword, Semantic & NER scoring. • Semantic scores enhance similarity matrix grading and impact evaluation. • NER scores address proper noun issues. • Deep Neural Network predicts actual marks effectively. • The best-performing LLM models are YAKE, SimCSE and Camembert. [ABSTRACT FROM AUTHOR]

The state-of-the-art performance on entity disambiguation has been reached by deep neural networks. However, the task remains very challenging due to the complexity of natural language. Moreover, the target data distribution is often different from that of training data. In this paper, we address the limitation of deep entity disambiguation from the perspective of misprediction risk. We propose a knowledge-based approach of risk analysis for entity disambiguation, and leverage it to enable adaptive deep learning. The proposed approach generates risk features by extracting evidences from the knowledge base, and then models them as a linearly-weighted random vector where an attention mechanism is used to focus on the most significant components. Finally, it estimates misprediction risk of the aggregated probability distribution via the Conditional Value-at-Risk metric. Furthermore, we demonstrate how to utilize risk analysis results in adaptive deep learning via two-phase training, the first phase fits on labeled training data while the second one minimizes misprediction risk on unlabeled target data. We evaluate the performance of the proposed approach on benchmark datasets through a comparative study. Our thorough experiments demonstrate that it can detect mispredictions more accurately than existing alternatives and can substantially improve the performance of deep learning models. [ABSTRACT FROM AUTHOR]

The burner combustion tuning is a complex problem that has been studied through flame monitoring and characterization. It has been observed that the flame electromagnetic spectrum and flickering contain specific flame information in combustion processes. This information is helpful for combustion stoichiometry tuning on burners. This paper described a method for selecting the best flame feature subset that can be computed from the scanner signal, in order to get the flame index and induce combustion stoichiometry on burners under specific combustion conditions. We propose a method for selecting a reduced subset with only the useful flame features for flame index classification. To extract the most relevant flame features we use a feature subset selection (FSS) algorithm and to determine the combustion state in burners, five flame indices were defined that represent the most common flame states in oil fuel-fired boilers. FSS includes complete, sequential, and random searches in order to eliminate redundant and noisy flame features to decrease the flame feature set dimension. A probabilistic neural network (PNN) algorithm was implemented for flame feature clustering. Signals from the actual flame scanner system and relevant variables from the boiler data acquisition system were used by the algorithms to calculate the burner flame index. A set of parametric tests was done in a heavy oil-fired boiler under well-known flame and index conditions to train and test the flame classifier. The results showed that only the four more relevant features are enough to classify flames with a good performance (92.3% accuracy), which is useful for burner combustion monitoring and optimization. • A new approach to defining/selecting critical features in the combustion process. • A new approach to classifying flames in industrial heavy oil-fired boilers. • The proposed algorithms are tested & evaluated in a real industrial process. • The overall methodology ensures that the combustion process guarantees safety. [ABSTRACT FROM AUTHOR]

Simulation-driven structural health monitoring (SHM) is crucial for enhancing the safety and reliability of structures. Its primary aim is to utilize numerical simulations to design efficient SHM systems, reducing costs while maintaining effectiveness. However, challenges arise due to disparities between simulations and physical measurements, stemming from uncertainties, noise, and simplifications. These challenges also extend to SHM methods using machine learning, particularly deep neural networks (DNNs). This paper bridges the gap between DNNs trained on simulated data and real-world experiments by investigating three transfer learning (TL) techniques: feature augmentation, direct fine-tuning, and boosting-based fine-tuning. The focus is on guided wave damage interaction coefficients (WDICs) as damage-sensitive features, modeled by DNNs for damage identification in thin plates, commonly found in aerospace structures like aircraft fuselages. This classification task compares measured and predicted WDICs in a common guided wave SHM configuration with one actuator and three sensors. Training and extensive testing datasets are generated using advanced finite element simulations and laser Doppler vibrometer experiments. The study demonstrates that TL techniques significantly enhance DNN prediction accuracy and improve damage identification performance for untrained scenarios compared to purely simulation-based DNNs. Among the TL approaches explored, feature augmentation with appropriate activation functions and boosting-based fine-tuning yield the most promising results. Using four out of twelve TL datasets, feature augmentation reduces the simulation-based mean absolute percentage error from 46.80% to 35.98% and increases coefficient of determination R 2 values from 0.02 to 0.72. With six TL datasets and a boosting-based fine tuning, the mean absolute percentage error decreases from 47.42% to 35.20%, maintaining similar R 2 values. While TL substantially enhances prediction accuracy, its impact on damage identification is slightly less pronounced, with improvements in relative classification accuracy by approximately 5% compared to simulation-based DNNs. These findings highlight the efficacy of TL techniques for damage identification in thin plates using WDICs and DNNs, emphasizing the influence of TL training and DNN model variations. Furthermore, they showcase the potential of TL methods to enhance DNN performance in future SHM applications, particularly emphasizing the promise of feature augmentation and boosting. • Novel simulation-driven structural health monitoring technique using guided waves. • Deep neural network modeling to predict wave damage interaction coefficients. • Bridging disparity between simulations and experiments through transfer learning. • Study of transfer learning techniques based on feature augmentation and fine tuning. • Demonstrated significant enhancement in simulation-based neural network performance. [ABSTRACT FROM AUTHOR]

Accurate and automatic segmentation of individual cell instances in microscopy images is a vital step for quantifying the cellular attributes, which can subsequently lead to new discoveries in biomedical research. In recent years, data-driven deep learning techniques have shown promising results in this task. Despite the success of these techniques, many fail to accurately segment cells in microscopy images with high cell density and low signal-to-noise ratio. In this paper, we propose a novel 3D cell segmentation approach DeepSeeded, a cascaded deep learning architecture that estimates seeds for a classical seeded watershed segmentation. The cascaded architecture enhances the cell interior and border information using Euclidean distance transforms and detects the cell seeds by performing voxel-wise classification. The data-driven seed estimation process proposed here allows segmenting touching cell instances from a dense, intensity-inhomogeneous microscopy image volume. We demonstrate the performance of the proposed method in segmenting 3D microscopy images of a particularly dense cell population called bacterial biofilms. Experimental results on synthetic and two real biofilm datasets suggest that the proposed method leads to superior segmentation results when compared to state-of-the-art deep learning methods and a classical method. • Data-driven seed estimation of a classical seeded watershed for 3D dense cell segmentation. • Enhancement of cell interior and border voxels by predicting Euclidean distance transforms using deep learning. • Multi-scale structural similarity index measure (MS-SSIM) loss for image-to- image mapping with better quality output. • Localization of cell seeds by performing voxel-wise classification using a deep network. [ABSTRACT FROM AUTHOR]

The aim of this paper is to explore the potential of the class of recurrent neural networks in developing a classification model for the stock market. For this research, the data that replicates the entire Nasdaq stock market limit order book is used. After extracting order book attributes from the raw dataset, feature selection approaches based on conditional entropy and Stochastic Universal Sampling are proposed, in order to highlight potentially informative features for our data classification task. Furthermore, the performances of nine presented recurrent neural network models for classifying the stock market data vector into one of the labels from the set S = {buy, sell, idle} are examined. In this study, the performances of the models based on different networks, namely the gated recurrent unit, the long short-term memory, and the recurrent neural network, are compared. Using a multi-criteria approach it is concluded that the model based on the GRU topology fed with the features selected using the newly proposed feature selection method outperforms other examined models. • The performances of nine different limit order book classification models are examined. • The PROMETHEE method specifies the most suitable model among the observed models. • The proposed feature demonstrates superiority compared to other selection methods. • The models based on different neural network topologies are evaluated. [ABSTRACT FROM AUTHOR]

The emotional states of speech have a great impact on the efficiency of speaker recognition (SR) system. Many researchers focus on how to map speech with different emotions to an emotion invariant embedding, which reduces the diversity of data. This paper proposes a new emotion embedding framework with self-attention mechanism for speaker recognition. First, several deep neural networks (DNNs) are trained to classify speakers in different emotional states as emotion embedding extractors during development phase. Then at enrollment stage, these pre-trained models are used to extend emotion embeddings from neutral speech. In order to make the final speaker embedding more representative, the classification model is trained with self-attention mechanism in emotion dimension, so that the framework can automatically annotate the weights of the emotion embeddings. Experiments were carried out on both Mandarin Affective Speech Corpus (MASC) and Crowd-Sourced Emotional Multimodal Actors Dataset (CREMA-D). The results show the proposed method achieves the best of Identification Rate (IR) and Equal Error Rate (EER) which are 59.14%, 15.79% on MASC and 75.98%, 8.14% on CREMA-D compared with state-of-the-art methods. In addition, the cross-database experiments also further demonstrate the practicability of the method in real scenes. • An emotion embedding framework for robust emotional speaker recognition is proposed. • Emotional feature extractors is pre-trained to obtain prior emotional representation. • Various emotion embeddings are extracted by decomposing the emotional features. • Self-attention is introduced to measure the importance of emotion embeddings. • Emotional speaker embedding enriches neutral speech with emotional information. [ABSTRACT FROM AUTHOR]

In this paper, we present a novel dimensionality reduction term (i.e. TripletPCA and ContrastivePCA) designed to enhance the robustness of pair-based loss functions in metric learning against image corruptions. Notably, our approach achieves this without the need for any data preprocessing steps. Our method can be seamlessly integrated into existing models, such as a block with Plug & Play framework. To the best of our knowledge, our TripletPCA and ContrastivePCA represent the first attempts to incorporate dimensionality reduction directly into pair-based metric learning losses for end-to-end metric learning. By projecting image features into low-dimensional vectors, our proposed loss functions effectively retain the essential components of images while mitigating the impact of corrupted features. Consequently, our metric learning loss function accurately computes feature distances through the projection. Experimental results demonstrate that our dimensionality reduction term can be easily incorporated into various types of existing deep neural networks. This integration leads to a substantial improvement in performance on standard benchmark datasets for corrupted image classification tasks. On several corruption datasets, we achieve an average performance improvement of 10.55% compared to existing baseline methods. Our code is available at https://github.com/Juryun/TDRL. [ABSTRACT FROM AUTHOR]

Spiking neural networks (SNNs) have garnered significant attention due to their ultra-high-speed and ultra-low-power operation, rendering them suitable for a range of energy-efficient applications. However, their performance in image generation is relatively mediocre, resulting in a lack of an effective generative replay mechanism to handle sequential probabilistic representations of multiple datasets. Consequently, this limitation often leads to catastrophic forgetting. This paper introduces a novel SNN lifelong learning framework namely Dynamic Lifelong learning with Spiking Generative Networks (DL-SGN), aimed at providing better generation ability while mitigating the problem of catastrophic forgetting in a continual learning scenario. DL-SGN comprises three key components: dynamic experts, a student, and an assistant. The dynamic expert is implemented as a dynamically expanding mixture model, with a proposed network expansion mechanism Dynamic Knowledge Adversarial Fusion (DKAF) facilitating the automatic handling of an increasing number of tasks. the student module adopts an SNN-based Variational Autoencoder (VAE) and classifier, leveraging accumulated knowledge from each expert. To enhance the student's ability to generalize to images, we introduce a discriminator as an assistant module trained using adversarial training. We validate the effectiveness of the proposed by conducting experiments on image generation and classification in a lifelong learning environment, and the results substantiate the effectiveness of DL-SGN. • The DL-SGN framework is proposed to addresses catastrophic forgetting in SNNs. • Our model is based on multiple dynamic experts and adversarial training. • Develop a dynamic knowledge adversarial fusion to expand expert module's capacity. • Employ an assistant module to enhance the student module's generalization ability. [ABSTRACT FROM AUTHOR]