River flow prediction is a pivotal task in the field of water resource management during the era of rapid climate change. The highly dynamic and evolving nature of the climatic variables e.g., precipitation has a significant impact on the temporal distribution of the river discharge in recent days making the discharge forecasting even more complicated for diversified water-related issues e.g., flood prediction and irrigation planning. To predict the discharge, various physics-based numerical models are used using numerous hydrologic parameters. Extensive lab-based investigation and calibration are required to reduce the uncertainty involved in those parameters. However, in the age of data-driven predictions, several deep learning algorithms showed satisfactory performance in dealing with sequential data. In this research, Long Short-term Memory (LSTM) neural network regression model is trained using over 80 years of daily data to forecast the discharge time series up to 3 days ahead of time. The performance of the model is found satisfactory through the comparison of the predicted data with the observed data, visualization of the distribution of the errors and Root Mean Squared Error (RMSE) value of 0.09. Higher performance is achieved through the increase in the number of epochs and hyper parameter tuning. This model can be transferred to other locations with proper feature engineering and optimization to perform univariate predictive analysis and potentially be used to perform real-time river discharge prediction.

Diabetic retinopathy, a common complication of diabetes, is further exacerbated by factors such as hypertension and obesity. This study introduces the Diabetic Retinopathy Convolutional Transformer (DRCT) model, which combines convolutional and transformer techniques to enhance the classification of retinal images. The DRCT model achieved an impressive average F1 score of 0.97, reflecting its high accuracy in detecting true positives while minimizing false positives. Throughout 100 training epochs, the model exhibited strong generalization capabilities, achieving superior validation accuracy with minimal overfitting. On a newly evaluated dataset, the model attained precision and recall scores of 96.93% and 98.89%, respectively, indicating a well-balanced handling of false positives and false negatives. The model's ability to classify retinal images into five distinct diabetic retinopathy categories demonstrates its potential to significantly improve automated diagnosis and aid in clinical decision-making.

Numerous approaches address over-fitting in neural networks: by imposing a penalty on the parameters of the network (L1, L2, etc); by changing the network stochastically (drop-out, Gaussian noise, etc.); or by transforming the input data (batch normalization, etc.). In contrast, we aim to ensure that a minimum amount of supporting evidence is present when fitting the model parameters to the training data. This, at the single neuron level, is equivalent to ensuring that both sides of the separating hyperplane (for a standard artificial neuron) have a minimum number of data points — noting that these points need not belong to the same class for the inner layers. We firstly benchmark the results of this approach on the standard Fashion-MINST dataset, comparing it to various regularization techniques. Interestingly, we note that by nudging each neuron to divide, at least in part, its input data, the resulting networks make use of each neuron, avoiding a hyperplane completely on one side of its input data (which is equivalent to a constant into the next layers). To illustrate this point, we study the prevalence of saturated nodes throughout training, showing that neurons are activated more frequently and earlier in training when using this regularization approach. A direct consequence of the improved neuron activation is that deep networks are now easier to train. This is crucially important when the network topology is not known a priori and fitting often remains stuck in a suboptimal local minima. We demonstrate this property by training a network of increasing depth (and constant width): most regularization approaches will result in increasingly frequent training failures (over different random seeds) whilst the proposed evidence-based regularization significantly outperforms in its ability to train deep networks.

This study aimed to explore the feasibility of using airborne acoustic emission in laser beam butt welding for the development of an automated classification system based on neural networks. The focus was on monitoring the formation of joint gaps during the welding process. To simulate various sizes of butt joint gaps, controlled welding experiments were conducted, and the emitted acoustic signals were captured using audible to ultrasonic microphones. To implement an automated monitoring system, a method based on short-time Fourier transformation was developed to extract audio features, and a convolutional neural network architecture with data augmentation was utilized. The results demonstrated that this non-destructive and non-invasive approach was highly effective in detecting joint gap formations, achieving an accuracy of 98%. Furthermore, the system exhibited promising potential for low latency monitoring of the welding process. The classification accuracy for various gap sizes reached up to 90%, providing valuable insights for characterizing and categorizing joint gaps accurately. Additionally, increasing the quantity of training data with quality annotations could potentially improve the classifier model's performance further. This suggests that there is room for future enhancements in the study.

Although commercially-available low-cost air quality sensors have low accuracy, the sensor system are being used to collect the data for the regulation of PM2.5 emission caused by industrial activities or to estimate the personal exposure for PM2.5. In this work, to solve the accuracy problem of low-cost PM sensor, we developed a new PM2.5 calibration model by combining the deep neural network (DNN) optimized in calibration problem and a LSTM optimized in time-dependent characteristics. First, two datasets were generated to test the accuracy performance and generalization performance of the PM2.5 calibration machine learning (ML) model. The PM2.5 concentrations, temperature and humidity by low-cost sensor and gravimetric-based PM2.5 measuring instrument were sampled for a sufficiently long time. The proposed model was compared with benchmark (multiple linear regression model) and low-cost sensor results. For root mean square error (RMSE) for PM2.5 concentrations, the proposed model reduced 41-60% of error compared to the raw data of low-cost sensor, and reduced 30-51% of error compared to the benchmark model. R2 of ML model, MLR and raw data were 93, 80 and 59 %. Also, the developed model still showed consistent calibration performance when calibrated with new sensors in different locations. Low-cost sensors combined with ML model not only can improve the calibration performance of benchmark, but also can be applied to the sensor monitoring systems for various epidemiologic investigations and regulatory decisions.

: Submerged aquatic vegetation (SAV) are highly efficient at carbon sequestration and, despite their relatively small distribution globally, are recognized as a potentially valuable component of climate change mitigation. However, SAV mapping in tidal marshes presents a challenge due to optically complex constituents in the water. The emergence and advancement of deep learn-ing-based techniques in the field of habitat mapping with remote sensing imagery provides an opportunity to address this challenge. In this study, an analytical framework was developed to quantify the carbon sequestration of SAV habitats in the Atchafalaya River Delta Estuary from field and remote sensing observations using deep convolutional neural network (DCNN) tech-niques. A U-Net based model, Wetland-SAV Network, was trained to identify SAV percent cover (high, medium, and low) as well as other estuarine habitat types from Landsat 8/9-OLI data. The areal extent of SAV was up to 8% of the total area (47,000 ha) with a significant loss of SAV habitats observed post-Hurricane Barry (~2,300 ha) in 2019. The habitat areas and habitat-specific carbon fluxes were then used to quantify net greenhouse gas (GHG) flux of the study area for with/without SAV scenarios in a Carbon Balance Model. The total net GHG flux was in the range of -0.13 ± 0.06 to -0.86 ± 0.37 ×105 tonne CO2e yr-1 and increased up to 40% (-0.23 ± 0.10 to -0.90 ± 0.39 ×105 tonne CO2e yr-1 ) when SAV was accounted for within the calculation. At the hectare scale, inclusion of SAV resulted in an increase of ~60% for net GHG sink in shallow areas adjacent to emergent marsh where SAV was abundant. This is the first attempt at remotely mapping SAV in coastal Louisiana as well as a first quantification of net GHG flux at the scale of hectares to thousands of hectares, accounting for SAV within these sub-tropical coastal delta marshes. Remote sensing and deep learning models have high potential for mapping and monitoring of SAV in turbid sub-tropical coastal deltas as a component of increasing accuracy of net GHG flux estimates at small (hectare) and large (coastal basin) scales.

Modern energy systems collect high volumes of data that can provide valuable information about energy consumption. Electric companies can now use historical data to make informed decisions on energy production by forecasting the expected demand. Many deep learning models have been proposed to deal with these type of time series forecasting problems. Deep neural networks, such as recurrent or convolutional, can automatically capture complex patterns in time series data and provide accurate predictions. In particular, Temporal Convolutional Networks (TCN) are a specialised architecture that has advantages over recurrent networks for forecasting tasks. TCNs are able to extract long-term patterns using dilated causal convolutions and residual blocks, and can also be more efficient in terms of computation time. In this work, we propose a TCN-based deep learning model to improve the predictive performance in energy demand forecasting. Two energy-related time series with data from Spain have been studied: the national electric demand, and the power demand at charging stations for electric vehicles. An extensive experimental study has been conducted, involving more than 1900 models with different architectures and parametrisations. The TCN proposal outperforms the forecasting accuracy of Long Short-Term Memory (LSTM) recurrent networks, which are considered the state-of-the-art in the field.

In response to the growing inspection demand exerted by process automation in component manufacturing, Non-destructive testing (NDT) continues to explore automated approaches that utilize deep learning algorithms for defect identification, including within digital X-ray radiography images. This necessitates a thorough understanding of the implication of image quality parameters on the performance of these deep learning models. This study investigates the influence of two image quality parameters, namely Signal-to-Noise Ratio (SNR) and Contrast-to-Noise Ratio (CNR), on the performance of U-net deep learning segmentation model. Input images were acquired with varying combinations of exposure factors such as kilovoltage, milli-ampere, and exposure time, which altered the resultant quality. The data was sorted into 5 different datasets according to their measured SNR and CNR values. The deep learning model was trained 5 distinct times, utilizing a unique dataset for each training session. Training the model with high CNR values yielded an intersection over Union (IoU) metric of 0.9594 on test data of the same category but drops to 0.5875 when tested on lower CNR test data. The result in this study emphasizes the importance of achieving a balance in training dataset according to the investigated quality parameters, to enhance the performance of deep learning segmentation models in NDT radiography applications.

Generative artificial intelligence (GenAI) has been developing with many incredible achievements like ChatGPT and Bard. Deep generative model (DGM) is a branch of GenAI, which is preeminent in generating raster data such as image and sound due to strong points of deep neural network (DNN) in inference and recognition. The built-in inference mechanism of DNN, which simulates and aims to synaptic plasticity of human neuron network, fosters generation ability of DGM which produces surprised results with support of statistical flexibility. Two popular approaches in DGM are Variational Autoencoders (VAE) and Generative Adversarial Network (GAN). Both VAE and GAN have their own strong points although they share and imply underline theory of statistics as well as incredible complex via hidden layers of DNN when DNN becomes effective encoding/decoding functions without concrete specifications. In this research, VAE and GAN is unified into a consistent and consolidated model called Adversarial Variational Autoencoders (AVA) in which VAE and GAN complement each other, for instance, VAE is good at generator by encoding data via excellent ideology of Kullback-Leibler divergence and GAN is a significantly important method to assess reliability of data which is realistic or fake. In other words, AVA aims to improve accuracy of generative models, besides AVA extends function of simple generative models. In methodology this research focuses on combination of applied mathematical concepts and skillful techniques of computer programming in order to implement and solve complicated problems as simply as possible.

With the increasing availability of large digitized fine art collections, automated analysis and classification of paintings is becoming an interesting area of research. However, due to domain specificity, implicit subjectivity, and pervasive nuances that vaguely separate art movements, analyzing art using machine learning techniques poses significant challenges. Residual networks, or variants thereof, are one the most popular tools for image classification tasks, which can extract relevant features for well-defined classes. In this case study, we focus on the classification of a selected painting 'Portrait of the Painter Charles Bruni' by Johann Kupetzky and the analysis of the performance of the proposed classifier. We show that the features extracted during residual network training can be useful for image retrieval within search systems in online art collections.

Tail-biting convolutional codes extend the classical zero-termination convolutional codes: Both encoding schemes force the equality of start and end states, but under the tail-biting each state is a valid termination. This paper proposes a machine-learning approach to improve the state-of-the-art decoding of tail-biting codes, focusing on the widely employed short length regime as in the LTE standard. This standard also includes a CRC code. First, we parameterize the circular Viterbi algorithm, a baseline decoder that exploits the circular nature of the underlying trellis. An ensemble combines multiple such weighted decoders, each decoder specializes in decoding words from a specific region of the channel words' distribution. A region corresponds to a subset of termination states; the ensemble covers the entire states space. A non-learnable gating satisfies two goals: it filters easily decoded words and mitigates the overhead of executing multiple weighted decoders. The CRC criterion is employed to choose only a subset of experts for decoding purpose. Our method achieves FER improvement of up to 0.75dB over the CVA in the waterfall region for multiple code lengths, adding negligible computational complexity compared to the circular Viterbi algorithm in high SNRs.

This paper reviews non-intrusive load monitoring (NILM) approaches that employ deep neural networks to disaggregate appliances from low frequency data, i.e. data with sampling rates lower than the AC base frequency. We first review the many degrees of freedom of these approaches, what has already been done in literature, and compile the main characteristics of the reviewed publications in an extensive overview table. The second part of the paper discusses selected aspects of the literature and corresponding research gaps. In particular, we do a performance comparison with respect to reported MAE and F$_1$-scores and observe different recurring elements in the best performing approaches, namely data sampling intervals below 10\,s, a large field of view, the usage of GAN losses, multi-task learning, and post-processing. Subsequently, multiple input features, multi-task learning and related research gaps are discussed, the need for comparative studies is highlighted, and finally, missing elements for a successful deployment of NILM approaches based on deep neural networks are pointed out. We conclude the review with an outlook on possible future scenarios.

Neural decoding from spiking activity is an essential tool for understanding the information encoded in population neurons, especially in applications like brain-computer interface (BCI). Various quantitative methods have been proposed and have shown superiorities under different scenarios respectively. From the machine learning perspective, the decoding task is to map the high-dimensional spatial & temporal neuronal activity to the low-dimensional physical quantities (e.g., velocity, position). Because of the complex interactions and the abundant dynamics among neural circuits, good decoding algorithms usually have the capability of capturing flexible spatiotemporal structures embedded in the input feature space. Recently, the Transformer-based models are widely used in processing natural languages and images due to its superior performances in handling long-range and global dependencies. Hence, in this work we examine the potential applications of Transformers in neural decoding and introduce two Transformer-based models. Besides adapting the Transformer to neuronal data, we also propose a data augmentation method for overcoming the data shortage issue. We test our models on three experimental datasets and their performances are comparable to the previous state-of-the-art (SOTA) RNN-based methods. In addition, Transformer-based models show increased decoding performances when the input sequences are longer, while LSTM-based models deteriorate quickly. Our research suggests that Transformer-based models are important additions to the existing neural decoding solutions, especially for large datasets with long temporal dependencies.

Decision tree algorithms have been among the most popular algorithms for interpretable (transparent) machine learning since the early 1980s. On the other hand, deep learning methods have boosted the capacity of machine learning algorithms and are now being used for non-trivial applications in various applied domains. But training a fully-connected deep feed-forward network by gradient-descent backpropagation is slow and requires arbitrary choices regarding the number of hidden units and layers. In this paper, we propose near-optimal neural regression trees, intending to make it much faster than deep feed-forward networks and for which it is not essential to specify the number of hidden units in the hidden layers of the neural network in advance. The key idea is to construct a decision tree and then simulate the decision tree with a neural network. This work aims to build a mathematical formulation of neural trees and gain the complementary benefits of both sparse optimal decision trees and neural trees. We propose near-optimal sparse neural trees (NSNT) that is shown to be asymptotically consistent and robust in nature. Additionally, the proposed NSNT model obtain a fast rate of convergence which is near-optimal up to some logarithmic factor. We comprehensively benchmark the proposed method on a sample of 80 datasets (40 classification datasets and 40 regression datasets) from the UCI machine learning repository. We establish that the proposed method is likely to outperform the current state-of-the-art methods (random forest, XGBoost, optimal classification tree, and near-optimal nonlinear trees) for the majority of the datasets.

Diabetes mellitus is a popular life-threatening disease and patients may gradually have started suffering from other diabetes-causing diseases such as heart attacks, stroke, hypertension, blurry vision, blindness, foot ulcer, amputation, kidney damage and other organ failures before diagnosis. Early detection can help reduce the fatality of this disease. Deep learning models have proven very useful in disease detection and computer-aided diagnosis. In this work, we proposed a deep unsupervised machine learning model for early detection of diabetes using voting ensemble feature selection and deep belief neural networks (DBN). Dataset was obtained from an online repository containing responses of prediagnosed patients to direct questionnaires administered in Sylhet Diabetes Hospital in Sylhet, Bangladesh. The dataset was preprocessed and preprocessed. Features were reduced using the ensemble feature selector. The DBN model was pretrained and tuned to obtain optimal performance. The model was also compared with other models with no multiple hidden layers. The DBN performed at its relative best with F1-measure, precision and recall of 1.00, 0.92 and 1.00 respectively. We conclude that DBN is a useful tool for an unsupervised early prediction of Type II diabetes mellitus.

This study investigates the factors influencing the choice of online and offline shopping trips and their impacts on urban transportation, environment, and economy in Tehran, Iran. A questionnaire survey was conducted to collect data from 1,000 active e-commerce users who made successful orders in both online and offline services in the last 20 days of 2021 in areas 2 and 5 of Tehran. A deep neural network model was developed to estimate the type of shopping trip based on 10 indicators, such as age, gender, car ownership, delivery cost, product price, etc. The performance of the model was compared with three other algorithms: MLP, decision tree, and KNN. The results showed that the deep neural network model had the highest accuracy of 95.63%. The most important factors affecting the choice of shopping trips were delivery cost, delivery time, and product price. This study provides insights for transportation planners, e-commerce managers, and policymakers to design effective strategies for reducing transportation costs, pollutant emissions, urban traffic, and increasing user satisfaction and sustainable development.

Speech signals are degraded in real-life environments, product of background noise or other factors. The processing of such signals for voice recognition and voice analysis systems presents important challenges. One of the conditions that make adverse quality difficult to handle in those systems is reverberation, produced by sound wave reflections that travel from the source to the microphone in multiple directions.To enhance signals in such adverse conditions, several deep learning-based methods have been proposed and proven to be effective. Recently, recurrent neural networks, especially those with long and short-term memory (LSTM), have presented surprising results in tasks related to time-dependent processing of signals, such as speech. One of the most challenging aspects of LSTM networks is the high computational cost of the training procedure, which has limited extended experimentation in several cases. In this work, we present a proposal to evaluate the hybrid models of neural networks to learn different reverberation conditions without any previous information. The results show that some combination of LSTM and perceptron layers produce good results in comparison to those from pure LSTM networks, given a fixed number of layers. The evaluation has been made based on quality measurements of the signal's spectrum, training time of the networks and statistical validation of results. Results help to affirm the fact that hybrid networks represent an important solution for speech signal enhancement, with advantages in efficiency, but without a significan drop in quality.

Recurrent Neural Networks (RNNs) have significantly advanced the field of machine learning by enabling the effective processing of sequential data. This paper provides a comprehensive review of RNNs and their applications, highlighting advancements in architectures such as Long Short-Term Memory (LSTM) networks, Gated Recurrent Units (GRUs), Bidirectional LSTM (BiLSTM), and stacked LSTM. The study examines the application of RNNs in different domains, including natural language processing (NLP), speech recognition, financial time series forecasting, bioinformatics, autonomous vehicles, and anomaly detection. Additionally, the study discusses recent innovations, such as the integration of attention mechanisms and the development of hybrid models that combine RNNs with convolutional neural networks (CNNs) and transformer architectures. This review aims to provide machine learning researchers and practitioners with a comprehensive overview of the current state and future directions of RNN research.

Recent progress in microfabrication technique allowed the rapid development of neural implants. They are getting categorized as effective tools for clinical practice, especially to treat traumatic and neurodegenerative disorders. Microelectrode arrays already have been used in numerous neural interface devices. Basically, almost all neural implants have been developed based on BCI (Brain Computer Interface) system. When BCI system falls under invasive technique, it is referred as BMI or Brain Machine Interface. BMIs hold promises for neurorehabilitation of motor and sensory function, cognitive state evaluation and treatment of neurological chaos. A directed overview of the field of neural implants is discussed in this article. The aim of this review is to give a brief introduction of neural prosthetics as well as their exciting applications in treating neurological disorders and a deep discussion on their functionality are mentioned. BCI system and their different types, their functionality, their pros and cons, how other neural implants developed, and their present status have been covered. Different possibilities and possible future of deep brain stimulation (DBS), Neuralink, motor and sensory neural prosthetics are further discussed.

Sign language recognition technology can help people with hearing impairments to communicate with those who are hearing impaired. At present, with the rapid development of society, deep learning technology also provided certain technical support for sign language recognition work. In sign language recognition tasks, the use of traditional convolutional neural networks to extract spatio-temporal features from sign language videos suffers from insufficient feature extraction, resulting in low recognition rates. Nevertheless, video-based sign language datasets are very large and require a lot of computational resources for training and generalisation must be ensured, which poses a challenge for recognition. In this paper, we have presented a video-based sign language recognition method based on resnet and lstm. As the number of network layers increases, the ResNet network can effectively solve the granularity explosion problem and obtain better time series features. We use the Resnet convolutional network as the backbone model. At the initialisation stage, we obtain sign language features using ResNet; then, the learned feature space is used as the input of LSTM network to obtain long sequence features. The experimental results show that the accuracy of the above model is better than the mainstream model, and it can effectively extract the spatio-temporal features in sign language videos and improve the recognition rate of sign language actions.

Estimating the remaining useful life (RUL) of components is a crucial task to enhance the reliability, safety, productivity, and to reduce maintenance cost. In general, predicting the RUL of a component includes constructing a health indicator ( ) to infer the current condition of the component, and modelling the degradation process, to estimate the future behavior. Although many signal processing and data-driven based methods were proposed to construct the , most of the existing methods are based on manual feature extraction techniques, and need the prior knowledge of experts, or rely on a large amount of failure data. Therefore, in this study, a new data-driven method based on the convolutional autoencoder (CAE) is presented to construct the . For this purpose, the continuous wavelet transform (CWT) technique is used to convert the raw acquired vibrational signals into a two-dimensional image; then, the CAE model is trained by the healthy operation dataset. Finally, the Mahalanobis Distance (MD) between the healthy and failure stages is measured as the . The proposed method is tested on a benchmark bearing dataset and compared with several other traditional construction models. Experimental results indicate that the constructed exhibits a monotonically increasing degradation trend and has a good performance to detect incipient faults.

In this paper we propose a computational framework that merges mechanistic modeling with deep neural networks obeying the Systems Biology Markup Language (SBML) standard. Over the last 20 years, the systems biology community has developed a large number of mechanistic models in SBML that are currently stored in public databases. With the proposed framework, existing SBML mechanistic models may be upgraded to hybrid systems through the incorporation of deep neural networks into the model core, using a freely available python tool. The so-formed hybrid mechanistic/neural network models are trained with a deep learning algorithm based on the adaptive moment estimation method (ADAM), stochastic regularization and semidirect sensitivity equations. The trained hybrid models are encoded in SBML and stored back in model databases, where they can be further analyzed as regular SBML models. The application of this approach is illustrated with three well-known case studies: the threonine synthesis model in Escherichia coli, the P58IPK signal transduction model, and the Yeast glycolytic oscillations model. The proposed framework is expected to greatly facilitate the widespread use of hybrid modeling techniques for systems biology applications.

The administration of regional neuraxial analgesia for pain relief during labor is widely recognized as a safe and effective method involving medication delivery into the epidural or subarachnoid space in the lower back. This study proposes an innovative semantic image segmentation methodology emphasizing enhanced interpretability using convolutional Random Fourier Features and layer-wise weighted class-activation maps tailored explicitly for foot segmentation in regional analgesia monitoring. Namely, our contribution is twofold: i) a novel Random Fourier Features layer is introduced to deal with image data to enhance three well-known architectures (FCN, UNet, and ResUNet); ii) three novel quantitive measures are presented to evaluate the interpretability of a given deep learning model devoted to segmentation tasks. Our approach is rigorously evaluated on a demanding dataset of foot thermal images from pregnant women who received epidural anesthesia. Its small size and considerable variability characterize the dataset. Our validation results demonstrate that the proposed methodology not only achieves competitive foot segmentation performance but also significantly enhances the explainability of the process, rendering it well-suited for applications such as epidural insertion during labor.

Abstract: Image aesthetic assessment (IAA) with neural attention has made significant progress due to its effectiveness in object recognition. Current studies have shown that the features learned by convolutional neural networks (CNN) at different learning stages indicate meaningful information. The shallow feature contains the low-level information of images and the deep feature perceives the image semantics and themes. Inspired by this, we propose a visual enhancement network with feature fusion (FF-VEN). It consists of two sub-modules, the visual enhancement module (VE module) and the shallow and deep feature fusion module (SDFF module). The former uses an adaptive filter in the spatial domain to simulate human eyes according to the region of interest (ROI) extracted by neural feedback. The latter not only takes out the shallow feature and the deep feature by transverse connection, but also uses a feature fusion unit (FFU) to fuse the pooled features together with the aim of information contribution maximization. Experiments on standard AVA dataset and Photo.net dataset show the effectiveness of FF-VEN.

In the field of low power wireless networks, one of the techniques on which many researchers are putting their efforts is related to positioning methodologies such as fingerprinting in dense urban areas. This paper presents an experimental study aimed at quantifying the mean location estimation error in densely urbanized areas.Using a dataset made available by the University of Antwerp, a neural network was implemented with the aim of providing the position of the end-devices. In this way it was possible to measure the mean location estimation error in an area with high urban density. The results obtained show an accuracy in the localization of the end-device of less than 150 meters.This result would make it possible to use the fingerprint instead of alternative, energy consuming, methodologies such as GPS in IoT (Internet of Things) applications where battery life is the primary requirement to be met.

In any document, graphical elements like tables, figures, and formulas contain essential information. The processing and interpretation of such information require specialized algorithms. Off-the-shelf OCR components cannot process this information reliably. Therefore, an essential step in document analysis pipelines is to detect these graphical components. It leads to a high-level conceptual understanding of the documents that makes digitization of documents viable. Since the advent of deep learning, the performance of deep learning-based object detection has improved many folds. In this work, we outline and summarize the deep learning approaches for detecting graphical page objects in the document images. Therefore, we discuss the most relevant deep learning-based approaches and state-of-the-art graphical page object detection in document images. This work provides a comprehensive understanding of the current state-of-the-art and related challenges. Furthermore, we discuss leading datasets along with the quantitative evaluation. Moreover, it discusses briefly the promising directions that can be utilized for further improvements.

Urban scene point cloud pose significant challenges for registration due to its large data volume, similar scenarios and dynamic objects. In this paper, we propose PCRMLP, a model for urban scene point cloud registration that achieves comparable registration performance to prior learning-based methods. Compared to previous works which focus on extracting features and estimating correspondence, the model estimates the transformation implicitly from concrete instances. An instance-level urban scene representation method is introduced to extract instance descriptors via semantic segmentation and DBSCAN, which enable the model to obtain robust instance features, filter dynamic objects and estimate transformation in a more logical manner. Then a lightweight network consisting of MLPs is employed to obtain transformation in an encoder-decoder manner. We validate the approach on KITTI dataset. Experimental results demonstrate that PCRMLP can obtain a satisfactory coarse transformation from instance descriptors just in 0.0028s. With a subsequent ICP refinement module, the proposed method achieves higher registration accuracy and computational efficiency than prior learning-based works.

In this paper we address the understanding of the problem, of why a deep learning model decides that an individual is eligible for a loan or not. Here we propose a novel approach for inferring, which attributes matter the most, for making a decision in each specific individual case. Specifically we leverage concepts from neural attention to devise a novel feature wise attention mechanism. As we show, using real world datasets, our approach offers unique insights into the importance of various features, by producing a decision explanation for each specific loan case. At the same time, we observe that our novel mechanism, generates decisions which are much closer to the decisions generated by human experts, compared to the existent competitors.

With the rapid developments that have been made in computer vision and computer technology, semantic segmentation of high-resolution images has emerged as a mainstream and challenging task. The, segmentation of water areas remains a complex task due to the similar features between water areas or between water areas and other land objects. Meanwhile, since coastlines and riverbanks are typically presented in irregular shapes, it is difficult to obtain the categories of pixels at boundaries. This paper proposed the use of a Water-Land Boundary Attention Network (WLBANet) to improve the accuracy of pixels at boundaries between water and land. This network also reinforced the performance of contextual information extraction into distinguishing between water areas and other land objects. To prove the validity of this proposed network, WLBANet was performed on the Water Land Dataset (WLD), which is also by this paper, and which contains water areas and other land objects categories. The results demonstrates that the proposed method has significant effectiveness.

Hand gesture recognition has many valuable applications in engineering and health care. This study proposes a novel model which can accurately distinguish hand gestures using forearm muscles' surface electromyogram (sEMG) signals. A deep learning algorithm with hyper parameters impacting the final model’s accuracy and a convolutional neural network (CNN) were employed in the recognition stage. The number of convolutional layers, kernels per layer, and neurons in the dense layer were selected for optimization, while the remaining parameters, such as the learning rate, batch size, and number of epochs, were chosen based on trial and error and prior knowledge. The optimal values for the selected hyperparameters were obtained using a genetic algorithm to achieve maximum recognition accuracy. The UC2018 Dual-Myo database was used for training and testing the model based on EMG signals characterizing the activity of eight different hand gestures. The final structure of the model consisted of two convolutional layers with 131 and 28 kernels, a dense layer with 111 neurons, and a softmax layer with eight neurons. Upon optimizing the hyperparameters using the genetic algorithm, the accuracy of the proposed model increased from 91.86% to 96.4% at best and 95.3% on average in real-time applications and 99.6% in an offline mode. Future work is warranted towards improving the architecture and the computational cost.

Complementary Metal-Oxide-Semiconductor (CMOS) is a typical image sensor that has a wide range of applications. However, considering the limitations of the weather condition and hardware cost, it is hard to capture high-resolution images by CMOS sensor. Recently, Super-Resolution (SR) techniques for image restoration has been gaining attentions due to its excellent performance. Under the powerful learning ability, Generative Adversarial Networks (GANs) have been proved to achieve great success. In this paper, we propose the Advanced Generative Adversarial Networks (AGAN) to efficiently correct these issues; 1) we design a Laplacian pyramid framework as pre-trained module, which is beneficial to provide multi-scale features for our input. 2) at each feature block, a convolutional skip-connections network, which may contain some latent information, is significant for generative model to reconstruct a plausible-looking image; 3) considering that edge details usually play an important role in image generation, a novel perceptual loss function is defined to train and seek optimal parameters. It is effective to achieve excellent and compelling quality captured by CMOS sensor. Quantitative and qualitative evaluations have been demonstrated that our algorithm not only fully takes advantage of Convolutional Neural Networks (CNNs) to improve the image quality, but also performs better than previous GAN algorithms for super-resolution task.

Background: Genome-wide association studies (GWAS) have shed light on various complex diseases and traits, by detecting more than 400,000 associated genetic loci. This number is expected to drastically increase because of the use of novel artificial intelligence methods offering new ways to study the effects of variants. Deep learning using artificial neural networks (ANN) is a sub-field of artificial intelligence, which simulates how the human brain learns. We aimed at assessing the potential of deep learning in human genetic studies of Alzheimer’s Disease (AD) and how these compare to the traditional statistical methods used in GWAS, by simultaneously testing the two approaches on the same dataset, while discovering new genetic loci associated to AD. Methods: To address this aim, phenotypic and genome-wide SNP data from the UK Biobank was analysed on a binary outcome, AD diagnosis, in two different data balance options, of one-to-one and one-to-two case-control datasets, using 2,764 cases vs 2,764 controls and 5,528 controls respectively matched on gender, age, ethnicity, PC1-20 and genotyping array. Genetic data handling and GWAS were performed using PLINK, whereas neural networks were trained using GenNet, a new ANN tool, with the same datasets, separated into training (60%), validation (20%) and test (20%) sets. Neural network layers were determined using biological knowledge, by annotating SNPs to genes and genes to AD related pathways, using ANNOVAR annotations followed by GeneSCF and KEGG. Results: Significant associations were detected between four SNPs linked to two different genes and AD for the 1 to 1 case-control study design and six SNPs linked to four different genes for the 1 to 2 case-control study design by using PLINK. All identified regions have been previously associated to AD. GenNet identified twelve SNPs on seven genes to be associated with AD, all with biological plausibility, achieving an AUC of 0.80 when using three biologically determined layers and 0.73 when using two layers at the neural networks. No common top SNPs were identified between the machine learning and GWAS models. Conclusion: This is one of the first studies attempting to compare the traditional GWAS to more sophisticated state-of-art methods for understanding the genetic architecture of complex phenotypes using the same dataset. More systematic comparisons with such approaches on real data are needed to enable best practises for machine learning in the analysis of genome-wide genetic data.

This paper presents an improved SOC estimation method for lithium ion batteries in Electric Vehicles using Bayesian optimized feedforward network. This innovative bayesian optimized neural network method attempts to minimize a scalar objective function by extracting hyperpa-rameters (hidden neurons in both layers) using a surrogate model. Furthemore, the hyperparameters are built and data samples are trained and validated. The performance of the proposed deep learning neural network is evaluated. Two reasonable size data samples are ex-tracted from Panasonic 18650PF Li-ion Mendeley datasets that are used for training and valida-tion. RNN and LSTM neural network algorithms offer the common core property of retaining past information and/or hidden states for better SOC estimation. However, the feature of this pro-posed method is the inclusion of Bayesian optimization that chooses optimal double layer hidden neurons. Analysis of results shows that Bayesian optimized feedforward algorithm with average MAPE (0.20%) is the lowest and is the best selection compared with average MAPE for other five deep learning algorithms. In the last quarter of fuel gauge, where fuel anxiety is severe, feed-forward with Bayesian Optimization algorithm is still the best selection (with MAPE of 0.64%).

Modelling both long and short-term user interests from historical data is crucial for accurate recommendations. However, unifying these metrics across multiple application domains can be challenging, and existing approaches often rely on complex, intertwined models which can be difficult to interpret. To address this issue, we propose a lightweight, plug-and-play interest enhancement module that fuses interest vectors from two independent models. After analyzing the dataset, we identify deviations in the recommendation performance of long and short-term interest models. To compensate for these differences, we use feature enhancement and loss correction during training. In the fusion process, we explicitly split long-term interest features with longer duration into multiple local features. We then use a shared attention mechanism to fuse multiple local features with short-term interest features to obtain interaction features. To correct for bias between models, we introduce a comparison learning task that monitors the similarity between local features, short-term features, and interaction features. This adaptively reduces the distance between similar features. Our proposed module combines and compares multiple independent long-term and short-term interest models on multiple domain datasets. As a result, it not only accelerates the convergence of the models but also achieves outstanding performance in challenging recommendation scenarios.

Vehicle re-identification research under surveillance cameras has yielded impressive results. However, the challenge of Unmanned Aerial Vehicle (UAV)-based vehicle re-identification (ReID) presents a high degree of flexibility, mainly due to complicated shooting angles, occlusions, low discrimination of top-down features, and significant changes in vehicle scales. To address this, we propose a novel Dual Mixing Attention Network (DMANet) to extract discriminative features robust to variations in viewpoint. Specifically, we first present a plug-and-play Dual Mixing Attention Module (DMAM) to capture pixel-level pairwise relationships and channel dependencies, where DMAM is composed of Spatial Mixing Attention (SMA) and Channel Mixing Attention (CMA). First, the original feature is divided according to the dimensions of spatial and channel to obtain multiple subspaces. Then, a learnable weight is applied to capture the dependencies between local features in the mixture space. Finally, the features extracted from all subspaces are aggregated to promote their comprehensive feature interaction. Moreover, the DMAM can be readily inserted into backbone networks at any depth to improve vehicle discrimination. The experiments show that the proposed structure gains a better performance compared with the representative methods in the UAV-based vehicle ReID task. Our code and models will be publicly released.

Micro-expression (ME) is one of the key psychological stress reactions. It is a modest, spontaneous facial mechanism. ME has significant applicability in a variety of psychologically-related sectors because to its precision and unpredictability with regard to psychological manifestations. Nevertheless, the current Micro-expression recognition (MER) algorithms have poor accuracy and a limited quantity of ME data, and this study issue has not been thoroughly investigated. Therefore, we present an approach for deep learning based on a Spatio-temporal capsule network (STCP-Net). STCP-Net has four components: a jitter reduction module, a differential feature extraction module, an STCP module, and a fully linked layer. The first two modules are aimed to extract diversifying differential features more precisely and to limit the influence of head jitter. The STCP module is used to extract Spatio-temporal features layer by layer, taking the temporal and geographical connection between features into account. This research runs sufficient trials using the Leave One Subject Out (LOSO) methodology for cross-validation using the CASMEII dataset. The conclusion and analysis demonstrate that the algorithm is innovative and efficient.

In videos, the human's actions are of three-dimensional (3D) signals. These videos investigate the spatiotemporal knowledge of human behavior. The promising ability is investigated using 3D convolution neural networks (CNNs). The 3D CNNs have not yet achieved high output for their well-established two-dimensional (2D) equivalents in still photographs. Board 3D Convolutional Memory and Spatiotemporal fusion face training difficulty preventing 3D CNN from accomplishing remarkable evaluation. In this paper, we implement Hybrid Deep Learning Architecture that combines STIP and 3D CNN features to enhance the performance of 3D videos effectively. After implementation, the more detailed and deeper charting for training in each circle of space-time fusion. The training model further enhances the results after handling complicated evaluations of models. The video classification model is used in this implemented model. Intelligent 3D Network Protocol for Multimedia Data Classification using Deep Learning is introduced to further understand space-time association in human endeavors. In the implementation of the result, the well-known dataset, i.e., UCF101 to, evaluates the performance of the proposed hybrid technique. The results beat the proposed hybrid technique that substantially beats the initial 3D CNNs. The results are compared with state-of-the-art frameworks from literature for action recognition on UCF101 with an accuracy of 95%.

Recently deep cross-modal hashing networks have received increasing interests due to its superior query efficiency and low storage cost. However, most of existing methods concentrate less on hash representations learning part, which means the semantic information of data cannot be fully used. Furthermore, they may neglect the high-ranking relevance and consistency of hash codes. To solve these problems, we propose a Self-Attention and Adversary Guided Hashing Network (SAAGHN). Specifically, it employs self-attention mechanism in hash representations learning part to extract rich semantic relevance information. Meanwhile, in order to keep invariability of hash codes, adversarial learning is adopted in the hash codes learning part. In addition, to generate higher-ranking hash codes and avoid local minima early, a new batch semi-hard cosine triplet loss and a cosine quantization loss are proposed. Extensive experiments on two benchmark datasets have shown that SAAGHN outperforms other baselines and achieves the state-of-the-art performance.

With the continuous development of network technology, complex network systems generate massive unbalanced attack traffic. Due to the severe imbalance in the quantities of normal samples and attack samples, as well as among different types of attack samples, intrusion detection systems suffer from low detection rates for rare class attack data. In this paper, we propose a geometric synthetic minority oversampling technique based on optimized kernel density estimation algorithm. This method can generate diverse rare class attack data by learning the distribution of rare class attack data while maintaining similarity with the original sample features. Meanwhile, the balanced data is input to a feature extraction module built upon multiple denoising autoencoders, reducing information redundancy in high-dimensional data and improving the detection performance for unknown attacks. Subsequently, a soft voting ensemble learning technique is utilized for multi-class anomaly detection on the balanced and dimensionally reduced data. Finally, an intrusion detection system is constructed based on data preprocessing, imbalance handling, feature extraction, and anomaly detection modules, and validated on the NSL-KDD and N-BaIoT datasets. Comparative experiments with baseline models and other state-of-the-art methods demonstrate that the proposed system improves the detection rate of rare class attack data. Furthermore, it achieves a good overall detection rate on the Internet of Things dataset (N-BaIoT), indicating its strong applicability.

Accurate electricity price forecasting has become a substantial requirement since the liberalization of the electricity markets. Due to the challenging nature of the electricity prices, which includes high volatility, sharp price spikes and seasonality, various types of electricity price forecasting models still compete and can not outperform each other consistently. Neural Networks have been successfully used in machine learning problems and Recurrent Neural Networks (RNNs) have been proposed to address time-dependent learning problems. In particular, Long Short Term Memory and Gated Recurrent Units (GRU) are tailor-made for time series price estimation. In this paper, we propose to use Gated Recurrent Units as a new technique for electricity price forecasting. We have trained a variety of algorithms with rolling 3-year window and compared the results with the RNNs. In our experiments, 3-layered GRUs outperformed all other neural network structures and state of the art statistical techniques in a statistically significant manner in the Turkish day-ahead market.

Over the last several decades, large wildfires are increasingly common across the United States causing disproportionate impact on forest health and function, human well-being, and economy. Here, we examine the severity of large wildfires across the Contiguous United States over the past decade (2011-2020) using a wide array of meteorological, vegetational, and topographical features in the Deep Neural Network model. A total of 4,538 wildfire incidents were used in the analysis covering 87,305 square miles of burned area. We observed the highest number of large wildfires in California, Texas, and Idaho, with lightning causing 43 % of these incidents. Importantly, results indicate that the severity of wildfire occurrences is highly correlated with the climatological forcings, land cover, location, and elevation of the ecosystem. Overall, results may serve useful guide in managing landscapes under changing climate and disturbance regimes.

Several researchers have contemplated deep learning-based post-filters to increase the quality of statistical parametric speech synthesis, which perform a mapping of the synthetic speech to the natural speech, considering the different parameters separately and trying to reduce the gap between them. The Long Short-term Memory (LSTM) Neural Networks have been applied successfully in this purpose, but there are still many aspects to improve in the results and in the process itself. In this paper, we introduce a new pre-training approach for the LSTM, with the objective of enhancing the quality of the synthesized speech, particularly in the spectrum, in a more efficient manner. Our approach begins with an auto-associative training of one LSTM network, which is used as an initialization for the post-filters. We show the advantages of this initialization for the enhancing of the Mel-Frequency Cepstral parameters of synthetic speech. Results show that the initialization succeeds in achieving better results in enhancing the statistical parametric speech spectrum in most cases when compared to the common random initialization approach of the networks.

A continuous Deep Belief Network (cDBN) with two hidden layers is proposed in this paper, focusing on the problem of weak feature learning ability when dealing with continuous data. In cDBN, the input data is trained in an unsupervised way by using continuous version of transfer functions, the contrastive divergence is designed in hidden layer training process to raise convergence speed, an improved dropout strategy is then implemented in unsupervised training to realize features learning by de-cooperating between the units, and then the network is fine-tuned using back propagation algorithm. Besides, hyper-parameters are analysed through stability analysis to assure the network can find the optimal. Finally, the experiments on Lorenz chaos series, CATS benchmark and other real world like CO2 and waste water parameters forecasting show that cDBN has the advantage of higher accuracy, simpler structure and faster convergence speed than other methods.

The research of object classification and part segmentation is a hot topic in computer vision. A considerable number of studies have been carried out about deep learning on 3D point clouds. However,it is challenging to achieve effective feature learning due to sparsity of point clouds. Recently, a variety of Transformers have been adopted to improve point cloud processing and display great potential. Nevertheless, large numbers of Transformer layers tend to incur huge computational and memory costs. PointNet++ is one of the most influential neural architectures for point cloud understanding. Although the accuracy of PointNet++ has been largely surpassed by recent networks, this does not mean that PointNet++ has no potential.Thus, this paper offer two major contributions that significantly improve PointNet++ performance. Firstly, we introduce a novel contextual feature extraction (CFE) block that significantly enhances the feature extraction capabilities of PointNet++ networks. Secondly, to further enhance feature fusion, we seamlessly integrate a mutual learning (ML) block into the network architecture. By embedding these two innovative blocks within each layer of the network, we not only enrich the network's functionality but also impart it with greater robustness and adaptability. The specific experiments were conducted on the S3DIS (6-fold cross-validation) and Modelnet40 datasets with 86.5% and 92.7% accuracy, respectively, which proved that our model is comparable or even better than most existing methods for classification and segmentation tasks.

Hyperspectral images (HSI) contain abundant spectral information. Efficient extraction and utilization of this information for image classification remain prominent research topics. Previously, hyperspectral classification techniques primarily relied on statistical attributes and mathematical models of spectral data. Deep learning classification techniques have recently been extensively utilized for hyperspectral data classification, yielding promising outcomes. This study proposes a deep learning approach that uses polarization feature maps for classification. Initially, the polar coordinate transformation method is employed to convert the spectral information of all pixels in the image into spectral feature maps. Subsequently, the proposed Deep Context Feature Fusion Network (DCFF-NET) is utilized to classify these feature maps. The model is validated using three open-source hyperspectral datasets: Indian Pines, Pavia University, and Salinas. The experimental results indicated that DCFF-NET achieves excellent classification performance. Experimental results on three public HSI datasets demonstrate that the proposed method accurately recognizes different objects with an overall accuracy (OA) of 86.68%, 94.73%, and 95.14% based on the pixel method, and 98.15%, 99.86%, and 99.98% based on the pixel-patch method.

Amid escalating urbanization, devising rational commercial space layouts is a critical challenge. Leveraging machine learning, this study uses a Back-propagation (BP) neural network to optimize commercial spaces in Weinan City's central urban area. The results indicate an increased number of commercial facilities with a trend of multi-centered agglomeration and outward expansion. Based on these findings, we propose a strategic framework for rational commercial space development emphasizing aggregation centers, development axes, and spatial guidelines. This strategy provides valuable insights for urban planners in small and medium-sized cities in the Yellow River Basin and metropolitan areas, ultimately showcasing the power of machine learning in enhancing urban planning.

Recently, deep neural networks (DNNs) have been widely used in various fields such as autonomous vehicles and smart homes. Since these DNNs can be directly implemented on edge devices, they offer advantages such as real-time processing in low-power and low-bandwidth environments. However, deployment of DNNs in embedded systems, including edge devices, exposes them to threats such as fault injection attacks. In this paper, we introduce two methods to induce misclassifications by using clock and voltage glitch-based fault attacks in devices where DNN models are executed. As a result of experiments on a microcontroller implemented with a DNN for image classification, we show that glitch injection attacks can lead with high probability to serious misclassifications. Furthermore, we propose a countermeasure to glitch attacks on the softmax function, and confirm that this method is effective in preventing misclassifications.

Next-generation cancer and oncology research needs to take full advantage of the multi-modal structured, or graph, information, with the graph datatypes ranging from molecular structures to spatially resolved imaging and digital pathology to biological networks to knowledge graphs. Graph Neural Networks (GNNs) efficiently combine the graph structure representations with the high predictive performance of deep learning, especially on the large multi-modal datasets. In this review article, we survey the landscape of recent (2020-present) GNN applications in the context of cancer and oncology research, and delineate six currently predominant research areas. Subsequently, we identify the most promising directions for future research. We compare GNNs with graphical models and "non-structured" deep learning, and devise the guidelines for cancer and oncology researchers or physician-scientists asking the question of whether they should adopt the GNN methodology in their research pipelines.

Oscillometric blood pressure (BP) devices currently estimate a single point but do not identify fluctuations in BP or distinguish them from variations in response to physiological properties. In this paper, to analyze BP normality based on oscillometric measurements, we use statistical approaches including kurtosis, skewness, Kolmogorov-Smirnov, and correlation tests. Then, to mitigate uncertainties, we use a deep neural network (DNN) to determine the confidence limits (CLs) of BP measurements based on their normality. The proposed DNN regression model decreases the standard deviation of error (SDE) of the mean error (ME) and the mean absolute error (MAE) and reduces the uncertainty of the CLs and SDEs of the proposed technique. We validate the normality of the distribution of the BP estimation distribution which fits the Gaussian distribution very well. We use a rank test in the DNN regression model to demonstrate the independence of the artificial SBP and DBP estimations. First, we perform statistical tests to verify the normality of the BP measurements for individual subjects. The proposed methodology provides accurate BP estimations and reduces the uncertainties associated with the CLs and SDEs based on the DNN regression estimator.

Recent advancements in artificial intelligence (AI) and deep learning have revolutionized the field of protein engineering, particularly in the area of de novo protein design. This review article explores the impact of AI-driven approaches on protein design, with a specific focus on the role of structure prediction neural networks, such as AlphaFold and RoseTTAFold. The article discusses the paradigm shift brought about by these networks, which have enabled the design of proteins with unique structures and functions that are not found in nature. The review covers various aspects of AI-driven protein design, including the use of protein language models to harness evolutionary information, the development of de novo protein design workflows utilizing deep learning, and the application of generative models like RFdiffusion and RoseTTAFold All-Atom. The article also highlights the successes and applications of AI-driven protein design across diverse domains, such as enzyme engineering, antibody design, and vaccine development. Additionally, the review identifies current challenges and future directions in the field, emphasizing the need to address limitations in modeling conformational dynamics and designing proteins for in vivo functionality. The article concludes by underscoring the potential of AI-driven protein design to transform various aspects of science and technology, while also acknowledging the importance of interdisciplinary collaboration and the development of robust pipelines for the validation and optimization of designed proteins. Overall, this comprehensive review serves as a valuable resource for researchers and practitioners interested in understanding the current state and future prospects of AI-driven protein design.

In the last decade, unprecedented progress in the development of neural networks influenced dozens of different industries, among which are signal processing for the electroencephalography process (EEG). Electroencephalography, even though it appeared in the first half of the 20th century, to this day didn’t change the physical principles of operation. But the signal processing technique due to the use of neural networks progressed significantly in this area. Evidence for this can serve that for the past 5 years more than 1000 publications on the topic of using machine learning have been published in popular libraries. Many different models of neural networks complicate the process of understanding the real situation in this area. In this manuscript, we provided the most comprehensive overview of research where were used neural networks for EEG signal processing.

Due to the complex underground environment, pumping machines are prone to produce numerous failures. The indicator diagrams of faults are similar in a certain degree, which produces indistinguishable samples. As the samples increases, manual diagnosis becomes difficult, which decreases the accuracy of fault diagnosis. For accurately and quickly judging the fault type, we propose an improved adaptive activation function and apply it to five types of neural networks. The adaptive activation function improves the negative semi-axis slope of the ReLU activation function by combining the gated channel conversion unit to improve the performance of the deep learning model. The proposed adaptive activation function is compared with the traditional activation function through the fault diagnosis data set and the public data set. The results show that the activation function has better nonlinearity, can improve the generalization performance of deep learning model, the accuracy of fault diagnosis. In addition, the proposed adaptive activation function can also be well embedded in other neural networks.

The colossal depths of the deep neural network sometimes suffer from ineffective backpropagation of the gradients through all its depths. Whereas, The strong performance of shallower multilayer neural structures prove their ability to increase the gradient signals in the early stages of training which easily gets backpropagated for global loss corrections. Shallow neural structures are always a good starting point for encouraging the sturdy feature characteristics of the input. In this research, a shallow, deep neural structure called PrimeNet is proposed. PrimeNet is aimed to dynamically identify and encourage the quality visual indicators from the input to be used by the subsequent deep network layers and increase the gradient signals in the lower stages of the training pipeline. In addition to this, the layerwise training is performed with the help of locally generated errors which means the gradient is not backpropagated to previous layers, and the hidden layer weights are updated during the forward pass, making this structure a backpropagation free variant. PrimeNet has obtained state-of-the-art results on various image datasets, attaining the dual objective of (1) compact dynamic deep neural structure, which (2) eliminates the problem of backwards-locking. The PrimeNet unit is proposed as an alternative to traditional convolution and dense blocks for faster and memory-efficient training, outperforming previously reported results aimed at adaptive methods for parallel and multilayer deep neural systems.

The implementation of neural networks (NN) on edge devices enables local processing of wireless data but faces challenges such as high computational complexity and memory requirements when deep neural networks (DNN) are used. Shallow neural networks customized for specific problems are more efficient, requiring fewer resources, and resulting in a lower latency solution. An additional benefit of the smaller network size is that it is suitable for real-time processing on edge devices. The main concern with shallow neural networks is their accuracy performance compared to DNNs. In this paper, we demonstrate that a customized adaptive activation function (AAF) can meet the accuracy of a DNN. We designed an efficient FPGA implementation for a customized segmented spline curve neural network (SSCNN) structure to replace the traditional fixed activation function with an AAF. We compared our SSCNN with different neural network structures such as real-valued time delay neural network (RVTDNN), augmented real-valued time delay neural network (ARVTDNN), and deep neural networks with different parameters. Our proposed SSCNN implementation uses 40% fewer hardware resources and no Block RAMS compared to the DNN with similar accuracy. We experimentally validate this computationally efficient and memory-saving FPGA implementation of SSCNN for digital predistortion of RF power amplifiers using the AMD/Xilinx RFSoC ZCU111. The implemented solution uses less than 3% of the available resources. The solution also enables an increase of the clock frequency to 221.12 MHz, allowing the transmission of wide bandwidth signals.

As modern systems are becoming more complex their control strategy cannot longer relay only on measurement information that usually comes from probes but also from mathematical models. Those systems models can lead to unbearable computation times due to their own complexity, turning the control process non-viable, which leads to the implementation of surrogate models that enable to achieve estimates within acceptable time to take decisions. A control trained with Deep Reinforcement Learning algorithm, using a Physics-Informed Neural Network to obtain the temperature map on the following time step, replaces the need of running Direct Numerical Simulations. On this work we considered an 1D heat conduction problem, which temperature distribution feeds a control system to activate a heat source aiming to obtain a constant, previously defined, temperature value. With this approach, control training becomes much faster without the need of performing numerical simulations or laboratory measurements, as well the control is taken based on Neural Network enabling its implementation on simple processors to edge computing.

In this paper, we perform the neural cryptanalysis and comprehensively analyze five block ciphers, Data Encryption Standard (DES), Simplified DES (SDES), Advanced Encryption Standard (AES), Simplified AES (SAES), and SPECK. The vulnerability of the block ciphers is investigated in three different attacks, such as Encryption Emulation (EE), Plaintext Recovery (PR), Key Recovery (KR), and Ciphertext Classification (CC) attacks. For the plaintexts, randomly generated block-sized bit arrays and texts are used. The block ciphers apply different numbers of round functions in the encryption of the block-sized bit arrays, and they are investigated by using the deep learning models trained with different numbers of data in EE, PR, and KR attacks. Moreover, the block ciphers use two different text encryption methods, Word-based Text Encryption (WTE) and Sentence-based Text Encryption (STE), to encrypt the texts in various operation modes, and they are analyzed with deep learning models in EE, PR, and CC attacks. As a result, the block ciphers can be vulnerable to deep learning-based EE and PR attacks using a large amount of train data, and STE can improve the strength of the block ciphers, unlike WTE, which shows almost the same classification accuracy as the plaintexts, especially in CC attack. Additionally, when the keys that are the same as the plaintexts are used in encryption, the block ciphers can be perfectly broken in KR attack. Moreover, especially in KR attack, the RNN-based deep learning model shows higher average Bit Accuracy Probability (BAPavg) than the fully connected-based deep learning model, which is used more often in previous works of neural cryptanalysis. Furthermore, although the transformer-based deep learning model is a state-of-the-art model in Natural Language Processing (NLP), the RNN-based deep learning model is more suitable for CC attack and shows higher classification accuracy.

Infrared and visible image fusion technologies are used to characterize the same scene by diverse modalities. However, most existing deep learning-based fusion methods are designed as symmetric networks, which ignore the differences between modal images and lead to the source image information loss during feature extraction. In this paper, we propose a new fusion framework for the different characteristics of infrared and visible images. Specifically, we design a dual-stream asymmetric network with two different feature extraction networks to extract infrared and visible feature maps respectively. The transformer architecture is introduced in the infrared feature extraction branch, which can force the network to focus on the local features of infrared images while still obtaining their contextual information. And the visible feature extraction branch uses residual dense blocks to fully extract the rich background and texture detail information of visible images. In this way, it can provide better infrared targets and visible details for the fused image. Experimental results on multiple datasets indicate that DSA-Net outperforms state-of-the-art methods in both qualitative and quantitative evaluations. In addition, we also apply the fusion results to the target detection task, which indirectly demonstrates the fusion performances of our method.

Deep learning has achieved lots of successes in many fields, but when trainable sample are extremely limited, deep learning often under or overfitting to few samples. Meta-learning was proposed to solve difficulties in few-shot learning and fast adaptive areas. Meta-learner learns to remember some common knowledge by training on large scale tasks sampled from a certain data distribution to equip generalization when facing unseen new tasks. Due to the limitation of samples, most approaches only use shallow neural network to avoid overfitting and reduce the difficulty of training process, that causes the waste of many extra information when adapting to unseen tasks. Euclidean space-based gradient descent also make meta-learner's update inaccurate. These issues cause many meta-learning model hard to extract feature from samples and update network parameters. In this paper, we propose a novel method by using multi-stage joint training approach to post the bottleneck during adapting process. To accelerate adapt procedure, we also constraint network to Stiefel manifold, thus meta-learner could perform more stable gradient descent in limited steps. Experiment on mini-ImageNet shows that our method reaches better accuracy under 5-way 1-shot and 5-way 5-shot conditions.

Automated skin lesion analysis is one of the trending fields that has gained attention among the dermatologists and healthcare practitioners. Skin lesion restoration is an essential preprocessing step for lesion enhancements for accurate automated analysis and diagnosis. Digital hair removal is a non-invasive method for image enhancement by solving the hair-occlusion artefact in previously captured images. Several hair removal methods were proposed for skin delineation and removal. However, manual annotation is one of the main challenges that hinder the validation of these proposed methods on a large number of images or using benchmarking datasets for comparison purposes. In the presented work, we propose a realistic hair simulator based on context-aware image synthesis using image-to-image translation techniques via conditional adversarial generative networks for generation of different hair occlusions in skin images, along with the ground-truth mask for hair location. Besides, we explored using three loss functions including L1-norm, L2-norm and structural similarity index (SSIM) to maximise the synthesis quality. For quantitatively evaluate the realism of image synthesis, the t-SNE feature mapping and Bland-Altman test are employed as objective metrics. Experimental results show the superior performance of our proposed method compared to previous methods for hair synthesis with plausible colours and preserving the integrity of the lesion texture.

A common spinal condition, spondylolisthesis is the presence of a relative back or forth displacement between the upper and lower vertebra due to one vertebra being oriented away from the smooth curvature of a normal spine. Aging-related illnesses such as degenerative spondylolisthesis are especially burdensome on social welfare and health-care systems in an aging society, especially radiologists and clinical physicians. Therefore, we proposed a computer aided diagnosis algorithm, named LumbarNet, for vertebral slippage detection on clinical X-ray images. Collaborating with i) a P-grade, ii) a piecewise slope detection scheme, and iii) a dynamic shift detection routine, LumbarNet was thus specialized for analyzing complex structural patterns in lumbar spine X-ray images and outcompeted other U-Net based methods. Extensive experiments on lumbar spine X-ray images in standard clinical practices showed that LumbarNet achieved a mean intersection over union value of 0.88 in vertebral region detection and an accuracy of 88.83% in vertebral slippage detection.

The small satellite market continues to grow year after year. A compound annual growth rate of 17% is estimated during the period between 2020 and 2025. Low-cost satellites can send a vast amount of images to be post-processed at the ground to improve the quality and extract detailed information. In this domain lies the resolution enhancement task, where a low-resolution image is converted to a higher resolution automatically. Deep learning approaches to Super-Resolution (SR) reached the state-of-the-art in multiple benchmarks; however, most of them were studied in a single-frame fashion. With satellite imagery, multi-frame images can be obtained at different conditions giving the possibility to add more information per image and improve the final analysis. In this context, we developed and applied to the PROBA-V dataset of multi-frame satellite images a model that recently topped the European Space Agency’s Multi-frame Super Resolution (MFSR) competition. The model is based on proven methods that worked on 2D images tweaked to work on 3D: the Wide Activation Super Resolution (WDSR) family. We show that with a simple 3D CNN residual architecture with WDSR blocks and a frame permutation technique as data augmentation better scores can be achieved than with more complex models. Moreover, the model requires few hardware resources, both for training and evaluation, so it can be applied directly from a personal laptop.

In the age of social media, images come from unreliable sources that may change their content to fit a narrative. Automated detection of forged images is a complex task, especially considering recent technological advances in image manipulation software. However, in the existing literature, it is possible to identify two general approaches adopted by detection methods, active and passive. Active techniques preemptively act on an image by inserting structures before any manipulation is made that can verify its authenticity. In contrast, passive methods analyze the content of an image in search of traces of manipulation. In this way, this research proposes a novel approach to image manipulation detection, combining two passive methods through neural networks, creating a generalist approach capable of detecting with greater accuracy than the methods that compose it. Furthermore, this work used a combination of four datasets available in the literature for training and evaluation. After training, the merged approach obtained an accuracy of 89.59% in the set of validation images, significantly higher than the model trained with only unaltered images, which obtained 78.64%, and the two other models trained using images with a feature selection method applied to enhance inconsistencies that obtained 68.02% for Error-Level Analysis images and 50.70% for the method using Discrete Wavelet Transform. In addition to the performance improvement, the proposed approach’s accuracy variation was lower than the other models.

The analysis of microscopy images has always been an important yet time consuming process in in materials science. Convolutional Neural Networks (CNNs) have been very successfully used for a number of tasks, such as image segmentation. However, training a CNN requires a large amount of hand annotated data, which can be a problem for material science data. We present a procedure to generate synthetic data based on ad-hoc parametric data modelling for enhancing generalization of trained neural network models. Especially for situations where it is not possible to gather a lot of data, such an approach is beneficial and may enable to train a neural network reasonably. Furthermore, we show that targeted data generation by adaptively sampling the parameter space of the generative models gives superior results compared to generating random data points.

The task of human action recognition (HAR) based on skeleton data is a challenging yet crucial technique owing to its wide-ranging applications in numerous domains, including patient monitoring, security surveillance, and observation of human-machine interactions. While numerous algorithms have been proposed in an attempt to distinguish between a myriad of activities, most practical applications necessitate highly accurate detection of specific activity types. This study proposes a novel and highly accurate spatiotemporal graph autoencoder network for HAR based on skeleton data. Furthermore, an extensive investigation was conducted employing diverse modalities. To this end, a spatiotemporal graph autoencoder was constructed to automatically learn both spatial and temporal patterns from human skeleton datasets. The powerful graph convolutional network, designated as GA-GCN, developed in this study, notably outperforms the majority of existing state-of-the-art methods when evaluated on two common datasets, namely NTU RGB+D and NTU RGB+D 120. On the first dataset, the proposed approach achieved accuracies of 92.3\% and 96.8\% for the cross-subject and cross-view evaluations, respectively. On the more challenging NTU RGB+D 120 dataset, GA-GCN attained accuracies of 88.8\% and 90.4\% for the cross-subject and cross-set evaluations, respectively.

Deep neural networks (DNN) have successfully delivered a cutting-edge performance in several fields. With the broader deployment of DNN models on critical applications, the security of DNNs becomes an active and yet nascent area. Attacks against DNNs can have catastrophic results, according to recent studies. Poisoning attacks, including backdoor and Trojan attacks, are one of the growing threats against DNNs. Having a wide-angle view of these evolving threats is essential to better understand the security issues. In this regard, creating a semantic model and a knowledge graph for poisoning attacks can reveal the relationships between attacks across intricate data to enhance the security knowledge landscape. In this paper, we propose a DNN Poisoning Attacks Ontology (DNNPAO) that would enhance knowledge sharing and enable further advancements in the field. To do so, we have performed a systematic review of the relevant literature to identify the current state. We collected 28,469 papers from IEEE, ScienceDirect, Web of Science, and Scopus databases, and from these papers, 712 research papers were screened in a rigorous process, and 55 poisoning attacks in DNNs were identified and classified. We extracted a taxonomy of the poisoning attacks as a scheme to develop DNNPAO. Subsequently, we used DNNPAO as a framework to create a knowledge base. Our findings open new lines of research within the field of AI security.

Target recognition mainly includes three approaches: optical image-based, echo detection-based, and passive signal analysis-based methods. Among them, the passive signal-based method is closely integrated with practical applications due to its strong environmental adaptability. Based on passive radar signal analysis, we design an "end-to-end" model that cascades a noise estimation network with a recognition network to identify working modes in noise environment. The noise estimation network is implemented based on U-Net, which adopts a method of feature extraction and reconstruction to adaptively estimate the noise mapping level of the sample, which can help the recognition network to reduce noise interference. Focusing on the characteristics of radar signal, the recognition network is realized based on Multi-Scale Convolutional Attention Network (MSCANet). Firstly, the deep group convolution is used to isolate the channel interaction in the shallow network. Then, through the multi-scale convolution module, finer-grained features of the signal are extracted without increasing the complexity of the model. Finally, the self-attention mechanism is used to suppress the influence of low-correlation and negative-correlation channels and spaces. This method overcomes the problem that the conventional method is seriously disturbed by noise. We validated the proposed method in 81 kinds of noise environments, achieving an average accuracy of 94.65%. Additionally, we discussed the performance of six machine learning algorithms and four deep learning algorithms. Compared to these methods, proposed MSCANet achieved an accuracy improvement of approximately 17%. Our method demonstrates better generalization and robustness.

The number of connected IoT devices is increasing significantly due to their many benefits, including automation, improved efficiency and quality of life, and waste reducing. However, these devices have several vulnerabilities that have led to the rapid growth in the number of attacks. Therefore, several machine learning-based Intrusion Detection Systems (IDS) tools have been developed to detect intrusions and suspicious activity to and from a host (HIDS - Host IDS) or, in general, within the traffic of a network (NIDS - Network IDS). The proposed work performs a comparative analysis and an ablative study among recent machine learning-based NIDS to develop a benchmark of the different proposed strategies. The proposed work compares both shallow learning algorithms, such as Decision Trees, Random Forest, Naïve Bayes, Logistic Regression, XGBoost, and Support Vector Machine, and deep learning algorithms, such as DNN, CNN, and LSTM, whose approach is relatively new in the literature. The algorithms are evaluated on the KDD-99, NSL-KDD, UNSW-NB15, and IoT-23 datasets. The results show that the NIDS tools based on deep learning approaches achieve better performance in detecting network anomalies than shallow learning approaches.

Early detection of wildfires has been limited using the sun-synchronous orbit satellites due to their low temporal resolution and wildfires’ fast spread in the early stage. NOAA’s geostationary weather satellites GOES-R can acquire images every 15 minutes at 2km spatial resolution, and have been used for early fire detection. However, advanced processing algorithms are needed to provide timely and reliable detection of wildfires. In this research, a deep learning framework, based on Gated Recurrent Units (GRU), is proposed to detect wildfires at early stage using GOES-R dense time series data. GRU model maintains good performance on temporal modelling while keep a simple architecture, makes it suitable to efficiently process time-series data. 36 different wildfires in North and South America under the coverage of GOES-R satellites are selected to assess the effectiveness of the GRU method. The detection times based on GOES-R are compared with VIIRS active fire products at 375m resolution in NASA’s Fire Information for Resource Management System (FIRMS). The results show that GRU-based GOES-R detections of the wildfires are earlier than that of the VIIRS active fire products in most of the study areas. Also, results from proposed method offer more precise location on the active fire at early stage than GOES-R Active Fire Product in mid-latitude and low-latitude regions.

With the advancement of photoelectric technology and computer image processing technology, the visual measurement method based on point clouds is gradually applied to the 3D measurement of large workpieces. Point cloud registration is a key step in 3D measurement, and its registration accuracy directly affects the accuracy of 3D measurements. In this study, we designed a novel MPCR-Net for multiple partial point cloud registration networks. First, an ideal point cloud was extracted from the CAD model of the workpiece and was used as the global template. Next, a deep neural network was used to search for the corresponding point groups between each partial point cloud and the global template point cloud. Then, the rigid body transformation matrix was learned according to these correspondence point groups to realize the registration of each partial point cloud. Finally, the iterative closest point algorithm was used to optimize the registration results to obtain a final point cloud model of the workpiece. We conducted point cloud registration experiments on untrained models and actual workpieces, and by comparing them with existing point cloud registration methods, we verified that the MPCR-Net could improve the accuracy and robustness of the 3D point cloud registration.

Energy consumption is a critical issue in the energy sector, and recent events such as the global energy crisis, costs, the need to reduce greenhouse emissions, and extreme weather conditions have increased the demand for energy efficiency. Accurately predicting energy consumption is one of the key steps to addressing inefficiency in energy consumption, and its optimization. In this regard, accurate predictions of energy consumption would not only help to minimize wastage, but also to save cost. In this article, we propose intelligent models using an ensemble of Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) neural networks for predicting energy consumption in smart buildings. The proposed model outperforms other state-of-the-art deep learning models for predicting minute energy consumption, with a mean square error of 0.109. The proposed model also accurately captures latent trends in the data that other models struggle to capture. The results highlight the potential of using hybrid deep learning models for improved energy efficiency management in smart buildings.

Electrical power quality is one of the main elements in power generation systems. At the same time, it is one of the most significant challenges regarding stability and reliability. Due to different switching devices in this type of architecture, different kinds of power generators, and non-linear loads are used for different industrial processes. As a result of this, the need to classify and analyze Power quality disturbance (PQD) to prevent and analyze the degradation of the system reliability affected by the non-linear and non-stationary oscillatory nature. This paper presents A Novel Mul-titasking Deep Neural Network (MDL) for the Classification and Analysis of Multiple Electrical Disturbances. The characteristics are extracted with a specialized and adaptive methodology for non-stationary signals, Empirical Mode Decomposition (EMD). The methodology’s design, devel-opment, and various performance tests are carried out with 28 different difficulty levels, such as severity, disturbance duration time, and noise in the 20 dB to 60 dB signal range. MDL was devel-oped with a diverse data set in difficulty and noise, with a quantity of 4500 records of different samples of multiple electrical disturbances. The analysis and classification methodology has an average accuracy percentage of 95% with multiple disturbances. In addition, an average accuracy percentage of 90% in analyzing important signal aspects for studying electrical power quality such as crest factor, Per Unit voltage analysis, Short Term Flicker Perceptibility (Pst), and Total Harmonic Distortion (THD), among others.

The accurate prediction of molecular properties such as lipophilicity and aqueous solubility are of great importance and pose challenges in several stages of the drug discovery pipeline. Machine learning methods like graph-based neural networks (GNNs) have shown exceptionally good performance in predicting these properties. In this work, we introduce a novel GNN architecture, called directed edge graph isomorphism network (D-GIN). It is composed of two distinct sub-architectures (D-MPNN, GIN) and achieves an improvement in accuracy over its sub-architectures employing various learning, and featurization strategies. We argue that combining models with different key aspects help make graph neural networks deeper and simultaneously increase their predictive power. Furthermore, we address current limitations in assessment of deep-learning models, namely, comparison of single training run performance metrics, and offer a more robust solution.

Accurate climate data at fine spatial resolution are essential for scientific research and the development and planning of crucial social systems, such as energy and agriculture. Among them, sea surface temperature plays a critical role as the associated El Niño-Southern Oscillation (ENSO) is considered a significant signal of global interannual climate system. In this paper, we propose an implicit neural representation-based interpolation method with temporal information (T_INRI) to reconstruct climate data of high spatial resolution, with sea surface temperature as the research object. Traditional deep learning models for generating high-resolution climate data are only applicable to fixed resolution enhancement scales. In contrast, the proposed T_INRI method is not limited to the enhancement scale provided during the training process and its results indicate that it can enhance low-resolution input by arbitrary scale. Additionally, we discuss the impact of temporal information on the generation of high-resolution climate data, specifically, which month the low-resolution sea surface temperature data is from. Our experimental results indicate that T_INRI is advantageous over traditional interpolation methods under different enhancement scales, and the temporal information can improve T_INRI performance for a different calendar month. We also examined the potential capability of T_INRI in recovering missing grid value. These results demonstrate that the proposed T_INRI is a promising method for generating high-resolution climate data and has significant implications for climate research and related applications.

The study of Internet of Vehicle (IoV) based on Fog Computing (FC) and Artificial Intelligent (AI) has attracted more and more attention. However, there are still many problems require to investigate such as time-critical, scalability, load-balancing, energy consumption, and so on. Focusing on these problems, we proposed an AI-based Vehicle-to-Everything (V2X) model for tasks and resource offloading model for IoVs network, which ensures reliable low-latency communication efficient tasks offloading in IoV network by using Software Defined Vehicular based FC (SDV-F) architecture. To fit to time-critical data transmission task distribution, the proposed model reduces unnecessary task allocation at the fog computing layer by proposing an AI-based task-allocation algorithm in IoV layer to implement task allocation of each vehicle. By applying AI technologies of Reinforcement Learning (RL), Markov decision process, and Deep Learning (DL), the proposed model intelligently makes decision on maximizing resource utilization at the fog layer, and minimizing the average end-to-end delay of time-critical IoV applications. The experiment demonstrates the proposed model can efficiently distribute the fog layer tasks while minimizing the delay.

Photogrammetry involves aerial photography of the earth’s surface and subsequently processing the images to provide a more accurate depiction of the area (Orthophotography). It’s used by the Spanish Instituto Geográfico Nacional to update road cartography but requires a significant amount of manual labor due to the need to perform visual inspection of all tiled images. Deep Learning techniques (artificial neural networks with more than one hidden layer) can perform road detection but it is still unclear how to find the optimal network architecture. Our system applies grammar guided genetic programming to the search of deep neural network architectures. In this kind of evolutive algorithm all the population individuals (here candidate network architectures) are constrained to rules specified by a grammar that defines valid and useful structural patterns to guide the search process. Grammar used includes well-known complex structures (e.g. Inception-like modules) combined with a custom designed mutation operator (dynamically links the mutation probability to structural diversity). Pilot results show that the system is able to design models for road detection that obtain test accuracies similar to that reached by state of the art models when evaluated over a dataset from the Spanish National Aerial Orthophotography Plan.

Infrared small target detection (ISTD) plays a crucial role in both civilian and military applications. The detection of small targets against dense cluttered backgrounds remains a challenging task, requiring the collaboration of false alarm source elimination and target detection. Existing approaches primarily focus on modeling targets while often neglecting false alarm sources. To address this limitation, we propose a Target and False Alarm Collaborative Detection Network to leverage the information provided by false alarm sources and the background. Firstly, we introduce a False Alarm Source Estimation Block (FEB) that estimates potential interferences present in the background by extracting features on multiple scales and employing stepwise upsampling for feature fusion. Subsequently, we propose a framework that employs multiple FEBs to eliminate false alarm sources across multiple scales. Finally, a Target Segmentation Block (TSB) is introduced to accurately segment the targets to produce the final detection result. Experiments on public datasets demonstrate that, compared to other methods, our model achieves high accuracy in segmenting targets while being able to extract false alarm sources which can be used for further studies.

In recent years, researchers have paid increasing attention on hyperspectral image (HSI) classification using deep learning methods. To improve the accuracy and reduce the training samples, we propose a double-branch dual-attention mechanism network (DBDA) for HSI classification in this paper. Two branches are designed in DBDA to capture plenty of spectral and spatial features contained in HSI. Furthermore, a channel attention block and a spatial attention block are applied to these two branches respectively, which enables DBDA to refine and optimize the extracted feature maps. A series of experiments on four hyperspectral datasets show that the proposed framework has superior performance to the state-of-the-art algorithm, especially when the training samples are signally lacking.

In this paper, we introduce ResNetMF, a groundbreaking approach that harnesses the power of residual network matrix factorization to revolutionize recommendation systems. ResNetMF integrates residual networks, renowned for their ability to capture intricate patterns and features, with matrix factorization techniques that excel in modelling user-item interactions. This fusion presents a novel solution that surpasses the limitations of traditional recommendation systems. Through comprehensive experimentation and evaluation of diverse datasets, ResNetMF demonstrates remarkable enhancements in recommendation accuracy and efficiency. By effectively capturing both linear and nonlinear relationships in user-item interactions, ResNetMF provides superior recommendation quality. The outcomes from experiments unequivocally highlight the superiority of ResNetMF over existing state-of-the-art recommendation approaches, thereby validating its innovative nature and underscoring its potential to shape the future of recommendation systems. Through the integration of the deep residual network, ResNetMF approach facilitates the training of neural networks, enabling them to explore the underlying data layers more comprehensively. Extensive experimentation and evaluation across various datasets provide compelling evidence for the superiority of ResNetMF. Moreover, the proposed method utilized natural language processing (NLP) techniques for targeted information dissemination in recommendation systems, emphasizing the importance of personalized and relevant recommendations for user satisfaction and engagement.

In recent years, different deep learning frameworks were introduced for hyperspectral image (HSI) classification. However, the proposed network models have a higher model complexity and do not provide high classification accuracy if few-shot learning is used. This paper pre-sents an HSI classification method that combines random patches network (RPNet) and re-cursive filtering (RF) to obtain informative deep features. The proposed method first convolves image bands with random patches to extract multi-level deep RPNet features. Thereafter, the RPNet feature set is subjected to dimension reduction through principal component analysis (PCA) and the extracted components are filtered using the RF procedure. Finally, HSI spectral features and the obtained RPNet-RF features are combined to classify the HSI using a support vector machine (SVM) classifier. In order to test the performance of the proposed RPNet-RF method, some experiments were performed on three widely known datasets using a few training samples for each class and classification results were compared with those obtained by other advanced HSI classification methods adopted for small training samples. The comparison showed that the RPNet-RF classification is characterized by higher values of such evaluation metrics as overall accuracy and Kappa coefficient (https://github.com/UchaevD/RPNet-RF).

Non-invasive, electroencephalography (EEG)-based brain-computer interfaces (BCIs) on motor imagery movements translate the subject’s motor intention into control signals through classifying the EEG patterns caused by different imagination tasks, e.g. hand movements. This type of BCI has been widely studied and used as an alternative mode of communication and environmental control for disabled patients, such as those suffering from a brainstem stroke or a spinal cord injury (SCI). Notwithstanding the success of traditional machine learning methods in classifying EEG signals, these methods still rely on hand-crafted features. The extraction of such features is a difficult task due to the high non-stationarity of EEG signals, which is a major cause for the stagnating progress in classification performance. Remarkable advances in deep learning methods allow end-to-end learning without any feature engineering, which could benefit BCI motor imagery applications. We developed three deep learning models: 1) a long short-term memory (LSTM); 2) a proposed spectrogram-based convolutional neural network model (pCNN); and 3) a recurrent convolutional neural network (RCNN), for decoding motor imagery movements directly from raw EEG signals without (manual) feature engineering. Results were evaluated on our own, publicly available, EEG data collected from 20 subjects and on an existing dataset known as 2b EEG dataset from "BCI Competition IV". Overall, better classification performance was achieved with deep learning models compared to state-of-the art machine learning techniques, which could chart a route ahead for developing new robust techniques for EEG signal decoding. We underpin this point by demonstrating the successful real-time control of a robotic arm using our CNN based BCI.

The main purpose of this article is to identify brain diseases such as Alzheimer's and Parkinson's in an intelligent way using Python software and processing the function of neurons in the neural network areas of the brain, which are classified into different components, and the data is based on the information obtained from FMRI images using algorithms artificial intelligence can intelligently diagnose the type of disease. Therefore, by using spatial clustering based on the deep learning method with noise algorithm (DBSCAN), fMRI brain image and fMRI brain surface complexity are classified, and then image preprocessing will be applied to fMRI images by using artificial intelligence neural networks (ANN) algorithm. became. Then, by using the ANN algorithm and Deep CNN topology of the brain neural network, the type of disease in the brain can be diagnosed by simulating the performance of an artificial neural network (ANN) in Python software. Deep machine learning-based methods, especially algorithms based on neural networks, have shown great success in medical image analysis for various tasks including brain tumor detection, segmentation, and classification, and Alzheimer's and epilepsy. In this research, we propose an integrated model based on a convolutional neural network for simultaneous recognition and classification of brain diseases for intelligent diagnosis of tumors and epilepsy. The average accuracy obtained from our proposed model is 97.89 percent accuracy and 91.66 percent specificity.

With the proliferation of electric vehicles (EVs) and the consequential increase in EV battery circulation, the need for accurate assessment of battery health and remaining useful life (RUL) is paramount, driven by environmentally friendly and sustainability goals. This study addresses this pressing concern by employing data-driven methods, specifically harnessing deep learning techniques to enhance RUL estimation for lithium-ion batteries (LIB). Leveraging the Toyota Research Institute Dataset, consisting of 124 lithium-ion batteries cycled to failure and encompassing key metrics such as capacity, temperature, resistance, and discharge time, our analysis substantially improves RUL prediction accuracy. Notably, the Convolutional-LSTM-Deep Neural Network (CLDNN) model and the Transformer-LSTM (Temporal-Transformer) model have emerged as standout Remaining Useful Life (RUL) predictors. The CLDNN model, in particular, achieved a remarkable Mean Absolute Error (MAE) of 84.012 and a Mean Absolute Percentage Error (MAPE) of 25.676. Similarly, the Temporal-Transformer model exhibited notable performance with an MAE of 85.134 and a MAPE of 28.7932. These impressive results were achieved through the application of Bayesian hyperparameter optimization, further enhancing the accuracy of predictive methods. These models were benchmarked against existing approaches, demonstrating superior results with an improvement in MAPE ranging from 4.01% to 7.12%.

This paper presents a comprehensive investigation into the applicability and performance of two prominent growth models, namely the Verhulst model and the Montroll model, in the context of modeling tumor cell growth dynamics. Leveraging the power of Physics-Informed Neural Networks (PINNs), we aim to assess and compare the predictive capabilities of these models against experimental data obtained from the growth patterns of tumor cells. We employ a dataset comprising detailed measurements of tumor cell growth to train and evaluate the Verhulst and Montroll models. By integrating PINNs, we not only account for experimental noise but also embed physical insights into the learning process, enabling the models to capture the underlying mechanisms governing tumor cell growth. Our findings reveal the strengths and limitations of each growth model in accurately representing tumor cell proliferation dynamics. Furthermore, the study sheds light on the impact of incorporating physics-informed constraints on the model predictions. The insights gained from this comparative analysis contribute to advancing our understanding of growth models and their applications in predicting complex biological phenomena, particularly in the realm of tumor cell proliferation.

This work presents an approach for detecting mathematical formulas in scanned document images. The proposed approach is end-to-end trainable. Since many OCR engines cannot reliably work with the formulas, it is essential to isolate them to obtain the clean text for information extraction from the document. Our proposed pipeline comprises a hybrid task cascade network with deformable convolutions and a Resnext101 backbone. Both of these modifications help in better detection. We evaluate the proposed approaches on the ICDAR-2017 POD and Marmot datasets and achieve an overall accuracy of 96% for the ICDAR-2017 POD dataset. We achieve an overall reduction of error of 13%. Furthermore, the results on Marmot datasets are improved for the isolated and embedded formulas. We achieved an accuracy of 98.78% for the isolated formula and 90.21% overall accuracy for embedded formulas. Consequently, it results in an error reduction rate of 43% for isolated and 17.9% for embedded formulas.

This comprehensive review explores the advancements in processing-in-memory (PIM) techniques for deep learning applications. It addresses the challenges faced by monolithic chip architectures and highlights the benefits of chiplet-based designs in terms of scalability, modularity, and flexibility. The review emphasizes the importance of dataflow-awareness, communication optimization, and thermal considerations in designing PIM-enabled manycore architectures. It discusses different machine learning workloads and their tailored dataflow requirements. Additionally, the review presents a heterogeneous PIM system for energy-efficient neural network training and discusses thermally efficient dataflow-aware monolithic 3D (M3D) NoC architectures for accelerating CNN inferencing. The advantages of TEFLON (Thermally Efficient Dataflow-Aware 3D NoC) over performance-optimized SFC-based counterparts are highlighted. Overall, this review provides valuable insights into the development and evaluation of chiplet and PIM architectures, emphasizing improved performance, energy efficiency, and inference accuracy in deep learning applications.

It is time-consuming for acquiring complete data by fully phase encoding in two orthogonal directions along with one frequency encoding direction. Undersampling in the 3D k-space is promising in accelerating such 3D MRI process. Though 3D undersampling can be conducted according to predefined probability density, the density based method is not optimal. Because of the large amount of 3D data and computational cost, it is challenging to perform data-driven and learning-based 3D undersampling and subsequent 3D reconstruction. To tackle this challenge, this paper proposes a deep neural network called EEUR-Net, realized by optimizing specific undersampling patterns for the fully-sampled 3D k-space data. Innovatively, our undersampling algorithm employs an end-to-end deep learning approach to optimize phase encoding patterns and uses a 3D U-Net for image reconstruction of undersampled data. Through end-to-end training, we obtain an optimized 3D undersampling pattern, which significantly enhances the quality of the reconstructed image under the same acceleration factor. A series of experiments on a knee MRI dataset demonstrated that, in comparison to standard random uniform, radial, Poisson and equispaced Cartesian undersampling schemes, our end-to-end learned undersampling pattern considerably improves the reconstruction quality of undersampled MRI images.

One of the critical problems in multiclass classification tasks is the imbalance of the dataset. This is especially true when using contemporary pre-trained neural networks, where, in fact, the last layers of the neural network are retrained. Therefore, the large datasets with highly unbalanced classes are not good for models’ training since the use of such a dataset leads to overfitting and, accordingly, poor metrics on test and validation datasets. In this paper the sensitivity to a dataset imbalance of Xception, ViT-384, ViT-224, VGG19, ResNet34, ResNet50, ResNet101, Inception_v3, DenseNet201, DenseNet161, DeIT was studied using a highly imbalanced dataset of 20,971 images sorted into 7 classes. It is shown that the best metrics were obtained when using a cropped dataset with augmentation of missing images in classes up to 15% of the initial number. So, the metrics can be increased by 2-6% compared to the metrics of the models on the initial unbalanced data set. Moreover, the metrics of the rare classes' classification also improved significantly – the TruePositive value can be increased by 0.3 and more. As result, the best approach to train considered networks on an initially unbalanced dataset was formulated.

This study discusses the application of deep learning technology in network intrusion detection systems (IDS) and focuses on a new model named CNN-Focal. First, through the review of traditional IDS technology, it analyzes its limitations in dealing with complex network traffic. Then, the design principle of the CNN-Focal model is described in detail, which uses threshold convolution and SoftMax multi-class classification technology to effectively improve abnormal traffic detection's accuracy and efficiency. The experimental results show that CNN-Focal performs well on the open data set, demonstrating the potential and advantages of its application in the natural network environment and providing a new perspective and method for further research of deep learning in the field of network security in the future.

In the context of smart agriculture, accurately estimating plant leaf chemical parameters is crucial for optimizing crop management and improving agricultural yield. Hyperspectral imaging, with its ability to capture detailed spectral information across various wavelengths, has emerged as a powerful tool in this regard. However, the complex and high-dimensional nature of hyperspectral data poses significant challenges in extracting meaningful features for precise estimation. To address this challenge, this study proposes an end-to-end estimation network for multiple chemical parameters of Astragalus leaves based on attention mechanism (AM) and multivariate hyperspectral features (AM-MHENet). We leveraging HybridSN and multilayer perception (MLP) to extract prominent features from the hyperspectral data of Astragalus membranaceus var. mongholicus (AMM) leaves and stems, as well as the surface and deep soil surrounding AMM roots. This methodology allows us to capture the most significant characteristics present in these hyperspectral data with high precision. The AM is subsequently used to assign weights and integrate the hyperspectral features extracted from different parts of the AMM. The MLP is then employed to simultaneously estimate the chlorophyll content (CC) and nitrogen content (NC) of AMM leaves. Compared with estimation networks that utilize only hyperspectral data from AMM leaves as input, our proposed end-to-end AM-MHENet demonstrates superior estimation performance. Specifically, AM-MHENet achieves an R2 of 0.983 with an RMSE of 0.73 for the estimation of CC in AMM leaves. For NC estimation, AM-MHENet achieves an R2 value of 0.977 with an RMSE of 0.27. These results underscore AM-MHENet’s effectiveness in significantly enhancing the accuracy of both CC and NC estimation in AMM leaves. Moreover, these findings indirectly suggest a strong correlation between the development of AMM leaves and stems, as well as the surface and deep soil surrounding the roots of AMM, and directly highlight the ability of AM to effectively focus on the relevant spectral features within the hyperspectral data. This findings from this study could offer valuable insights into the simultaneous estimation of multiple chemical parameters in plants, thereby making a contribution to the existing body of research in this field.

Malignant growth is the most widely recognized repulsive infections winning around the world, and the patients with disease are saved just when the malignant growth is distinguished at the beginning phase. Each kind of disease is interesting, with its own arrangement of development properties and hereditary changes. This paper presents the lung knob division and disease characterization by proposing an enhancement calculation. The general technique of the created approach includes four stages, such as pre-processing, division, highlight extraction, and the order. From the outset, the CT picture of the lung is taken care of to the division. When the division is done, the highlights are extricated through morphological and measurable and surface highlights like LOOP and LGP. At long last, the extricated highlights are given to the order step. Here, the characterization is done dependent on the Deep Belief Network (DBN) which is prepared by utilizing the proposed Chicken-Sine Cosine Algorithm (CSCA) which distinguish the lung tumor, giving two classes in particular, knob or non-knob. The presentation assessment of lung knob division and malignant growth grouping dependent on CSCA is figured utilizing three measurements to be specific, precision, affectability, and the explicitness.

This paper represents a groundbreaking paradigm shift in network optimization. Departing from traditional static methodologies, this innovative approach harnesses the power of Generative Artificial Intelligence (AI) and Large Language Models (LLMs) to optimize cloud networks dynamically. By integrating advanced AI algorithms, this framework continuously adapts and evolves, ensuring optimal real-time performance. This dynamic optimization enhances efficiency and resilience, allowing cloud networks to adjust seamlessly to changing demands and conditions. Through the fusion of cutting-edge technology and adaptive intelligence, this approach heralds a new era in network optimization, empowering organizations to achieve unprecedent-ed levels of agility and scalability in their cloud infrastructures.

With the popularity of solar energy in the electricity market, demand arises for data such as precise locations of solar panels for efficient energy planning, management, and distribution. However, this data is not easily accessible and in some cases, information such as precise locations does not exist. Furthermore, existing data sets for training semantic segmentation models of PV installations are limited, and their annotation is time-consuming and labor-intensive. Therefore, for additional remote sensing (RS) data creation, the pix2pix generative adversarial network (GAN) is utilized, enriching the original resampled training data of varying GSDs without compromising its integrity. Experiments done with the DeepLabV3 model, ResNet-50 backbone, and pix2pix GAN architecture were conducted to find the optimal configuration for an accurate RS imagery segmentation model. The result is a fine-tuned solar panel semantic segmentation model, trained using transfer learning and utilizing an optimal amount – 60% of generated RS imagery for additional training data, increasing model accuracy. The findings demonstrate the benefits of using GAN-generated images as additional training data, increasing the size of small data sets, and improving the capabilities of the segmentation model for solar panel detection in RS images.

This paper presents a novel approach for accurate counting and localization of tropical plants in aerial images that is able to work in new visual domains in which the available data is not labeled. Our approach uses deep learning and domain adaptation, designed to handle domain shift between the training and test data, which is a common challenge in this agricultural applications. This method uses a source dataset with annotated plants and a target dataset without annotations, and adapts a model trained on the source dataset to the target dataset using unsupervised domain alignment and pseudolabeling. The experimental results show the effectiveness of this approach for plant counting in aerial images of pineapples under significative domain shift, achieving a reduction up to 97% in the counting error when compared to the supervised baseline.

Recently, a layer-stacked ESN model named deep echo state Network (DeepESN) has been established. As an interactional model of recurrent neural network and deep neural network, investigations of DeepESN are of significant importance in both areas. Optimizing the structure of neural networks remains a common task in artificial neural networks, and the question of how many neurons should be used in each layer of DeepESN must be stressed. In this paper, our aim is to solve the problem of choosing the optimized size of DeepESN. Inspired by the sensitive iterative pruning algorithm, a neuronal similarity-based iterative pruning merging algorithm (NS-IPMA) is proposed to iteratively prune or merge the most similar neurons in DeepESN. Two chaotic time series prediction tasks are applied to demonstrate the effectiveness of NS-IPMA. The results show that the DeepESN pruned by NS-IPMA outperforms unpruned DeepESN with the same network size, and NS-IPMA is a feasible and superior approach to improving the generalization performance of DeepESN.

This paper presents a novel architecture for detecting mathematical formulas in document images, which is an important step for reliable information extraction in several domains. Recently, Cascade Mask R-CNN networks have been introduced to solve object detection in computer vision. In this paper, we suggest a couple of modifications to the existing Cascade Mask R-CNN architecture: First, the proposed network uses deformable convolutions instead of conventional convolutions in the backbone network to spot areas of interest better. Second, it uses a dual backbone of ResNeXt-101, having composite connections at the parallel stages. Finally, our proposed network is end-to-end trainable. We evaluate the proposed approach on the ICDAR-2017 POD and Marmot datasets. The proposed approach demonstrates state-of-the-art performance on ICDAR-2017 POD at a higher IoU threshold with an f1-score of 0.917, reducing the relative error by 7.8%. Moreover, we accomplished correct detection accuracy of 81.3% on embedded formulas on the Marmot dataset, which results in a relative error reduction of 30%.

One of the critical tools for early detection and subsequent evaluation of the incidence of lung diseases is chest radiography. At a time when the speed and reliability of results, especially for COVID-19 positive patients, is important, the development of applications that would facilitate the work of untrained staff involved in the evaluation is also crucial. Our model takes the form of a simple and intuitive application, into which you only need to upload X-rays: tens or hundreds at once. In just a few seconds, the physician will determine the patient's diagnosis, including the percentage accuracy of the estimate. While the original idea was a mere binary classifier that could tell if a patient was suffering from pneumonia or not, in this paper we present a model that distinguishes between a bacterial disease, a viral infection, or a finding caused by COVID-19. The aim of this research is to demonstrate whether pneumonia can be detected or even spatially localized using a uniform, supervised classification.

Underwater robotics equipped with visual detection systems have the function of detecting un-derwater artifacts, which is of great significance to deep-sea archaeology. Underwater artifacts are located in complex environments with poor imaging conditions, and the targets have chal-lenging problems such as breakage, stacking, and sediment burial, which causes the marine cul-tural target detection failures. To solve these problems, this paper proposes an underwater cul-tural target detection algorithm based on the deformable deep aggregation network model for underwater robotics exploration. To fully extract the target feature information of underwater targets in complex environments, this paper designs a multi-scale deep aggregation network with deformable convolutional layers. Besides, the BAM attention module is designed for fea-ture optimization, which enhances the potential feature information of the target while weak-ening the background interference information. Finally, the target prediction is achieved through feature fusion at different scales. The proposed algorithm has been extensively validat-ed and analyzed on the collected underwater artifact datasets, and the precision, recall, and mAP of the algorithm have reached 93.1%, 91.4%, and 92.8%, respectively. In addition, the present al-gorithm has been practically deployed on the underwater robotics. In the deep-sea tests, the ar-tifact detection frame rate reaches up to 19 fps, which satisfies the real-time target detection。

A real-time monitoring system for the operational status of fishing vessels is an essential element for the modernization of the fishing industry. The operational status of fishing vessels can be identified by using onboard cameras to detect the deck crew and the use of fishing nets. Due to the typically limited processing capacity of shipboard equipment and the significant memory consumption of detection models, however, general target detection models are unable to perform real-time image detection to identify the operational status of fishing vessels. In this paper, we propose a lightweight real-time deck crew and the use of fishing nets detection method, YOLOv5s-SGC. It is based on the YOLOv5s model, which uses surveillance cameras to obtain video of fishing vessels operating at sea and enhances the dataset. YOLOv5s-SGC replaces YOLOv5s’s backbone and the ordinary convolutional blocks in the feature fusion network with ShufflNetV2 and Ghost module.

The power converter is the significant device in a wind power system. Wind turbine will be shut down and off grid immediately with the occurrence of the IGBT module open-circuit fault of power converter, which will seriously impact the stability of grid and even threaten personal safety. However, in the existing diagnosis strategies of power converter, there are few single and double IGBT modules open-circuit fault diagnosis methods producing negative results including erroneous judgment, omissive judgment and low accuracy. In this paper, a novel method to diagnose the single and double IGBT modules open-circuit faults of the permanent magnet synchronous generator (PMSG) wind turbine grid-side converter (GSC) is proposed. Above all, collecting the three-phase current varying with wind speed of 22 failure states including a normal state of PMSG wind turbine GSC as the original signal data. Afterward, the original signal data are decomposed by using variational mode decomposition (VMD) to obtain the mode coefficient series, which are analyzed by the proposed method base on fault trend feature for extracting the trend feature vectors. Finally, the trend feature vectors are used as the input of deep belief network (DBN) for decision-making and obtaining the classification results. The simulation and experimental results show that the proposed method can diagnose the single and double IGBT modules open-circuit faults of GSC, and the accuracy is higher than the benchmark models.

This study investigates the application of deep residual networks for predicting the dynamics of interacting three-dimensional rigid bodies. We present a framework combining a 3D physics simulator implemented in C++ with a deep learning model constructed using PyTorch. The simulator generates training data encompassing linear and angular motion, elastic collisions, fluid friction, gravitational effects, and damping. Our deep residual network, consisting of an input layer, multiple residual blocks, and an output layer, is designed to handle the complexities of 3D dynamics. We evaluate the network's performance using a dataset of 10,000 simulated scenarios, each involving 3-5 interacting rigid bodies. The model achieves a mean squared error of 0.015 for position predictions and 0.022 for orientation predictions, representing a 25% improvement over baseline methods. Our results demonstrate the network's ability to capture intricate physical interactions, with particular success in predicting elastic collisions and rotational dynamics. This work significantly contributes to physics-informed machine learning by showcasing the immense potential of deep residual networks in modeling complex 3D physical systems. We discuss our approach's limitations and propose future directions for improving generalization to more diverse object shapes and materials.