Sigmoid Activation Function Research Articles

STNet: Low-Complexity Neural Network Decoder With Network Pruning

In this letter, a novel and compact neural network called Sigmoid-Tanh Network (STNet) is proposed for channel decoding, which is only composed of sigmoid and tanh activation functions. To address the structural redundancy problem in long and short-term memory network (LSTM), the neurons in the STNet are redesigned with the most effective structure in the LSTM cell for decoding. To further reduce the computational complexity, we propose an automatic pruning method based on multiple layer sensitivity, which can effectively remove redundant weights in STNet decoder with slight performance loss. Simulation results show that the proposed STNet decoder achieves near-maximum likelihood (ML) performance with only 17.1% trainable parameters compared to LSTM. Moreover, our pruning method achieves comparable decoding performance when reducing 58.3% Floating-point operations (FLOPs) for STNet.

Forecasting Indicators of Economic Development of Ukraine Using an Artificial Neural Network

The article presents the results of forecasting indicators that characterise the economic development of Ukraine with the help of an artificial neural network developed by the author. For this purpose, a fully connected artificial neural network of direct signal propagation with a sigmoidal activation function was used. An inverse error propagation algorithm was used to train the artificial neural network. Used to assess the quality of the neural network Mean Squared Error and Mean Absolute Percentage Error. According to the forecasting results for the period 2022–2023, we have: a slight decline in Ukraine's GDP (US dollars), the growing role of the private sector in the structure of fixed assets of the country, and investment growth and growth in export earnings. At the same time, it is possible to reduce the amount of investment by the state and reduce the average salary (US dollars). Within the forecast period, the number of employed people in Ukraine will remain stable

Improved Perceptron of Subsurface Chlorophyll Maxima by a Deep Neural Network: A Case Study with BGC-Argo Float Data in the Northwestern Pacific Ocean

Subsurface chlorophyll maxima (SCMs), commonly occurring beneath the surface mixed layer in coastal seas and open oceans, account for main changes in depth-integrated primary production and hence significantly contribute to the global carbon cycle. To fill the gap of previous methods (in situ measurement, remote sensing, and the extrapolating function based on surface-ocean data) for obtaining SCM characteristics (intensity, depth, and thickness), we developed an improved deep neural network (IDNN) model using a Gaussian radial basis activation function to retrieve the vertical profile of chlorophyll a concentration (Chl a) and associated SCM characteristics from surface-ocean data. The annually averaged SCM depth was further incorporated into the bias term and the Gaussian activation function to improve the estimation accuracy of the IDNN model. Based on the Biogeochemical-Argo (BGC-Argo) data acquired for three regions in the northwestern Pacific Ocean, vertical Chl a profiles produced by our improved DNN model using sea surface Chl a and sea surface temperature (SST) were in good agreement with the observations, especially in regions with low surface Chl a. Compared to other neural-network-based models with one hidden layer and a sigmoid activation function, the IDNN model retrieved vertical Chl a profiles well in more eutrophic subpolar regions. Furthermore, the application of the IDNN model to infer vertical Chl a profiles from remote-sensing information was validated in the northwestern Pacific Ocean.

Design morphology complexity and conceptual building project cost forecasting

PurposeThis research aims to develop conceptual phase building project cost forecasting models by exploring the relationship of existing plan shape complexity indices and general design morphology parameters with total construction cost.Design/methodology/approachPlan shape indices proposed to date by the literature for measuring building design complexity are critically reviewed. Building morphology is also dictated by town planning restrictions such as plot coverage ratio or number of storeys. This study analyses historical data collected from 49 residential building projects to develop multiple linear regression (MLR) and artificial neural network (ANN) models for forecasting construction cost. Existing plan shape coefficients are calculated to evaluate the geometrical complexity of sampled projects. Ten regression-based cost estimating equations are totally derived from stepwise backward and forward methods, and their predictive accuracy is contrasted: to performance levels reported in past studies and to ANN models developed in this research with multilayer perceptron architecture.FindingsAnalysis of plan shape indices revealed that 85.7% of examined past projects possess a high degree of design complexity, hence resulting in expensive initial decisions. This highlights the need for more effective early design stage decision-making by developing new building economic tools. The most accurate regression model, with a mean absolute percentage error (MAPE) of 19.2%, predicts the log of total cost from wall to floor index and total building envelope surface. Other explanatory variables resulting in MAPE values in the order of 20%–22% are total volume, volume above ground level, gross floor area below ground level, gross floor area per storey and total number of storeys. The overall MAPE of regression-based equations is 24.3% whilst ANN models are slightly more accurate with MAPE scores of 21.8% and 21.6% for one hidden and two hidden layers, respectively. The most accurate forecasting model in the research is the ANN with two hidden layers and the sigmoid activation function which predicts total building cost from total building volume (19.1%).Originality/valueThis paper introduces MLR-based and ANN-based conceptual construction cost forecasting models which are founded solely on building morphology design parameters and compare favourably with previous studies with an average predictive accuracy less than 25%. This paper is expected to be beneficial to both practitioners and academics in the built environment towards more effective cost planning of building projects. The methodology suggested can further be implemented in other countries provided that accurate and relevant data from historical projects are used.

Influence of input values on the prediction model error using artificial neural network based on Taguchi's orthogonal array

AbstractRapid and accurate assessment of software project development using artificial intelligence tools can be essential for success in the software industry. This article has two objectives: to reduce the magnitude relative error (MRE) value in estimating the effort and cost of software development using the proposed artificial neural network architecture based on the Taguchi method and examine the influence of input variables on the change in relative error value. Clustering and fuzzification methods further mitigate the heterogeneous structure of the different project values of the datasets used. Taguchi method contributes to the reduction of the number of iterations by 99%, which achieves a significant reduction in estimation and value of MRE. By monitoring additional criteria, such as prediction, correlation, and comparing two activation functions, such as sigmoid and radial basis function, the proposed model's correctness, reliability, and stability are confirmed. Significantly better results are expected using the sigmoid activation function and a decrease in the value of the mean (MRE).

Remote Sensing Image Recognition Based on LOG-T-SSA-LSSVM and AE-ELM Network.

Aiming at the influence of different working conditions on recognition accuracy in remote sensing image recognition, this paper adopts hierarchical strategy to construct a network. Firstly, in order to establish the classification relationship between different samples, labeled samples are used for classification. A Logistic-T-distribution-Sparrow Search Algorithm-Least Squares Support Vector Machines (LOG-T-SSA-LSSVM) classification network is proposed. LOG-T-SSA algorithm is used to optimize parameters in LSSVM to establish a better network to achieve accurate classification between sample sets and then identify according to different categories. Through UCI dataset test, the accuracy of LOG-T-SSA-LSSVM network classification is significantly improved compared with that of contrast network. The autoencoder is integrated with Extreme Learning Machine, and the autoencoder is used to realize data compression. The advantages of Extreme Learning Machine (ELM) network, such as less training parameters, fast learning speed, and strong generalization ability, are fully utilized to realize efficient and supervised recognition. Experiments verify that the autoencoder-extreme learning machine (AE-ELM) network has a good recognition effect when the sigmoid activation function is selected and the number of hidden layer neurons are 2000. Finally, after image recognition under different working conditions, it is proved that the recognition accuracy of AE-ELM based on LOG-T-SSA-LSSVM classification is significantly improved compared with traditional ELM network and Particle Swarm Optimization-Extreme Learning Machine (PSO-ELM) network.

Wound Detection by Simple Feedforward Neural Network

Chronic wounds are a heavy burden on medical facilities, so any help in treating them is most welcome. Current research focuses on wound analysis, especially wound tissue classification, wound measurement, and wound healing prediction to assist medical personnel in wound treatment, with the main goal of reducing wound healing time. The first phase of wound analysis is wound segmentation, where the task is to extract wounds from the healthy tissue and image background. In this work, a standard feedforward neural network was developed for the purpose of wound segmentation using data from the MICCAI 2021 Foot Ulcer Segmentation (FUSeg) Challenge. It proved to be a simple yet efficient method for extracting wounds from images. The proposed algorithm is part of a compact system that analyzes chronic wounds using a robotic manipulator, RGB-D camera and 3D scanner. The feedforward neural network consists of only five fully connected layers, the first four with Rectified Linear Unit (ReLU) activation functions and the last with sigmoid activation functions. Three separate models were trained and tested using images provided as part of the challenge. The predicted images were post-processed and merged to improve the final segmentation performance.The accuracy metrics observed during model training and selection were Precision, Recall and F1 score. The experimental results of the proposed network provided a recall value of 0.77, precision value of 0.72, and an F1 score (Dice score) of 0.74.

Random vector functional link with ε-insensitive Huber loss function for biomedical data classification

Background and objectiveBiomedical data classification has been a trending topic among researchers during the last decade. Biomedical datasets may contain several features noises. Hence, the conventional machine learning model cannot efficiently handle the presence of noise in datasets. Among the several machine learning model, the random vector functional link (RVFL) is one of the most popular and efficient models for task related to both classification and regression. Despite its excellent classification performance, its performance degrades while dealing with the datasets with noise. Researchers are searching for powerful models to minimize the influence of noise in datasets. Therefore, to enhance the classification ability of RVFL on noisy datasets, this paper suggests a novel random vector functional link with ε-insensitive Huber loss function (ε-HRVFL) for biomedical data classification problems. MethodsThe optimization problem of ε-HRVFL is reformulated as strongly convex minimization problems with a simple function iterative approach to find solutions. To have a better understanding of the scope of the biomedical data classification problem and potential solutions, we conducted experiments with three different types of label noise in biomedical datasets as well as a few non-biomedical datasets. The classification accuracy of the proposed ε-HRVFL model is compared statistically using Friedman test with the support vector machine, extreme learning machine with radial basis function (RBF) and sigmoid activation functions and RVFL with RBF and sigmoid activation functions. ResultsFor non-biomedical datasets, the proposed model showed the highest accuracy of 98.1332%. Moreover, for the biomedical datasets, the proposed model showed the best accuracy of 96.5229%. The proposed ε-HRVFL model with sigmoid activation function reveals the best mean ranks among the reported classifiers for both, biomedical and non-biomedical datasets. ConclusionNumerical results show the applicability of the proposed ε-HRVFL model. In future, the proposed ε-HRVFL can be developed to solve multiclass biomedical data classification problems. Moreover, ε-insensitive asymmetric Huber loss function based RVFL model can be developed for dealing more efficiently with these noisy biomedical datasets.

Spatial Attention Guided Residual Attention Network for Hyperspectral Image Classification

Hyperspectral image (HSI) classification has become a research hotspot. Recently, deep learning-based methods have achieved preferable performances by which the deep spectral-spatial features can be extracted from HSI cubes. However, in complex scenes, due to the diversity of the types of land-cover and the bands in high dimensional, these methods are often hampered by the irrelevant spatial areas and the redundant bands, which results in the indistinguishable features and the restricted performance. In this article, a spatial attention guided residual attention network (SpaAG-RAN) is proposed for HSI classification, which contains a spatial attention module (SpaAM), a spectral attention module (SpeAM), and a spectral-spatial feature extraction module (SSFEM). Based on the spectral similarity, the SpaAM is capable of capturing the relevant spatial areas composed of the pixels of the same category as the center pixel from HSI cube with a novel inverted-shifted-scaled sigmoid activation function. The SpeAM aims to select the bands which are beneficial to the spectral features representation. The SSFEM is exploited to extract the discriminating spectral-spatial features. To facilitate the processes of bands selection and features extraction, two well-designed spatial attention masks generated by the SpaAM are employed to guide the works of the SpeAM and the SSFEM, respectively. Moreover, a spatial consistency loss function is installed to maintain the consistency between the two spatial attention masks so that the network enables the distinction of the relevant features exactly. Experimental results on three HSI data sets show that the proposed SpaAG-RAN model can extract the discriminating spectral-spatial features and outperforms the state-of-the-arts.

SPECIFICS OF THE LEARNING ERROR DEPENDENCE OF MULTILAYERED NEURAL NETWORKS FROM THE ACTIVATION FUNCTION DURING THE PROCESS OF PRINTED DIGITS IDENTIFICATION

In this paper, we investigated the learning process during an identification of printed digits from the type of an activation function. The study of the activation function type and the number of iterations in the learning process of the neural system was carried out using the Fourier spectra analysis of the learning error function and branching diagrams. For this purpose, a program for the multilayer neural network was developed in the Python software, which involves setting the number of hidden layers as well as the number of neurons inside them and changes in the learning rate. The learning rate was considered as a constant, and its optimal value, where the best learning rate is observed, was determined. To analyze the learning rate effect on the educational process, we used a logistic function describing the frequency doubling process. It is shown that the learning error function is characterized by bifurcation processes leading to a chaotic state when η >0.8. The optimal learning rate value that determines the emergence of the doubling process of the local minima number is determined. It was found that the sigmoidal activation function (as compared to the activation functions ReLU and hyperbolic tangent) best satisfies the learning process of the three-layer neural network for recognizing digits, given an array of 4x7 zeros and ones. Compared to other activation functions, there is an insignificant change in the learning error during the transition from one digit to another. It is shown that an increase in the number of hidden layers does not lead to a sharp increase during the learning error. An increase in the learning iterations number is accompanied by the appearance of periodic dependences of the logistic function value of the learning rate, the period of which is a variable of the number of iterations and the learning rate. Using Fourier spectra of the error function from the learning rate value, it can be argued that an increase in the number of iterations leads to an increase in the number of harmonics, which eventually leads to the appearance of a chaotic state of the neural network. Keywords : Multilayered neural network, activation function, optimal learning rate, digital identification.

Historical Arabic Images Classification and Retrieval Using Siamese Deep Learning Model

Classifying the visual features in images to retrieve a specific image is a significant problem within the computer vision field especially when dealing with historical faded colored images. Thus, there were lots of efforts trying to automate the classification operation and retrieve similar images accurately. To reach this goal, we developed a VGG19 deep convolutional neural network to extract the visual features from the images automatically. Then, the distances among the extracted features vectors are measured and a similarity score is generated using a Siamese deep neural network. The Siamese model built and trained at first from scratch but, it didn't generated high evaluation metrices. Thus, we re-built it from VGG19 pre-trained deep learning model to generate higher evaluation metrices. Afterward, three different distance metrics combined with the Sigmoid activation function are experimented looking for the most accurate method for measuring the similarities among the retrieved images. Reaching that the highest evaluation parameters generated using the Cosine distance metric. Moreover, the Graphics Processing Unit (GPU) utilized to run the code instead of running it on the Central Processing Unit (CPU). This step optimized the execution further since it expedited both the training and the retrieval time efficiently. After extensive experimentation, we reached satisfactory solution recording 0.98 and 0.99 F-score for the classification and for the retrieval, respectively.

Symbolic Regression Based Feature Extraction of Shallow Neural-Networks for Identification and Prediction

This paper proposes a feature extraction method to improve the performance of shallow neural-network models with less number of parameters to apply especially on embedded system design at remote applications. Feature extraction method is designed using fuzzy c-means clustering based fuzzy system design cascaded a layer of symbolic operators and functions, respectively. During the training stage of neural-networks, symbolic operators and functions are selected using random-learning theory with the unity internal weights such that based on the prediction performance, optimal sequences are recorded for feature extraction to be utilized on testing phase. Extracted features are here used to empower the single-layer neural-network (SLNN) with sigmoid hyperbolic activation functions, functional-link neural- network (FLNN) with Chebyshev polynomials and Pi-Sigma higher-order neural-network (PSNN) with sigmoid activation functions, respectively. The internal and output parameters of the appended shallow neural-networks are optimized using batch optimization methods. Proposed regression models are first tested on identification of an artificial discrete-time dynamic system and real-time inverted pendulum then also for prediction of the sunspot time-series and traffic density estimation. As a result, the prediction performance of shallow neural networks is improved to be used in future applications.

Deep neural network algorithm for MPPT control of double diode equation based PV module

One of the most promising renewable energy sources (RES) is photovoltaic energy system. It generates power by using a photovoltaic (PV) module. The solar illumination and temperature value are the main parameters for the PV module because a small change can affect the entire power generation. The PV module is based on the double diode equation for extracting unknown parameter values and tracking maximum power point using basic analytical methods (Newton Raphson and Lambert W-function), particle swarm optimization (PSO) and artificial neural network (ANN) algorithms are performed. Then, a deep neural network (DNN) algorithm is proposed in this work for alleviating the disadvantages of the existing methods. Particularly, the DNN algorithm is proposed for reducing the mismatching power loss. Firstly, two-parameter values retrieve from a specific location at NASA open-access source for the year 2020. Then, the parameter values optimization through two analytical methods such as Newton Raphson (NR) and Lambert-W function (W-function) methods. Based on these analytical methods, the characteristic curve is executed. It is compared one among one and also the analytical methods evaluates with soft computing algorithms. The soft computing algorithm represents particle swarm optimization and the artificial neural network. Both the analytical methods and the soft computing algorithms compare for the extraction of the unknown parameter value for the PV module. Then, the existing algorithm is compared with advanced soft computing algorithms. In that deep neural network, the algorithm is implemented for the maximum power point tracking (MPPT) process. The DNN algorithm contains two inputs, three hidden layers and one output. The input neuron is defined with the voltage and current value and the hidden layer contains the sigmoid activation function. The result evaluates in terms of percentage of error for maximum power point (PEmpp).

Performance analysis of artificial neural network models for hour-ahead electric load forecasting

Supervised Artificial Neural Networks (ANN) is considered as a popular machine learning framework for year-ahead, month-ahead and day-ahead electric load forecasting but is yet to be optimized for very short term predictions like hour-ahead forecasting granularity. This study conducted a performance analysis of ANN models for hour-ahead electric load forecasting that power utility companies can use for agile reaction to its participation among spot markets. Data preparation procedures were conducted which transformed the historical electric load records of a certain geographic area served by a power utility into appropriate forms resulting to partitioned, represented and normalized datasets for ANN training and testing processes. Thirty-six ANN models all having nine input neurons and one output neuron were evaluated and found out that ANN models with Sigmoid activation function exhibited promising forecasting performance shown in terms of Mean Absolute Percentage Error (MAPE) with Back Propagation training algorithm with three hidden neurons having 2.85% MAPE; Quick Propagation training algorithm with six hidden neurons having 2.91% MAPE; and Resilient Propagation training algorithm with six hidden neurons having 3.49% MAPE. The performance analysis presented in this study shows how these ANN models were able to generate close to accurate forecasting results which power utility companies can effectively use for optimal resource management as they participate in spot markets that will require internationally accepted forecasting tolerance error in hour-ahead electric load nomination.

An Adaptive Sigmoidal Activation Function for Training Feed Forward Neural Network Equalizer

Feed for word neural networks (FFNN) have attracted a great attention, in digital communication area. Especially they are investigated as nonlinear equalizers at the receiver, to mitigate channel distortions and additive noise. The major drawback of the FFNN is their extensive training. We present a new approach to enhance their training efficiency by adapting the activation function. Adapting procedure for activation function extensively increases the flexibility and the nonlinear approximation capability of FFNN. Consequently, the learning process presents better performances, offers more flexibility and enhances nonlinear capability of NN structure thus the final state kept away from undesired saturation regions. The effectiveness of the proposed method is demonstrated through different challenging channel models, it performs quite well for nonlinear channels which are severe and hard to equalize. The performance is measured throughout, convergence properties, minimum bit error achieved. The proposed algorithm was found to converge rapidly, and accomplish the minimum steady state value. All simulation shows that the proposed method improves significantly the training efficiency of FFNN based equalizer compared to the standard training one.

Inverse Result of Approximation for the Max-Product Neural Network Operators of the Kantorovich Type and Their Saturation Order

In this paper, we consider the max-product neural network operators of the Kantorovich type based on certain linear combinations of sigmoidal and ReLU activation functions. In general, it is well-known that max-product type operators have applications in problems related to probability and fuzzy theory, involving both real and interval/set valued functions. In particular, here we face inverse approximation problems for the above family of sub-linear operators. We first establish their saturation order for a certain class of functions; i.e., we show that if a continuous and non-decreasing function f can be approximated by a rate of convergence higher than 1/n, as n goes to +∞, then f must be a constant. Furthermore, we prove a local inverse theorem of approximation; i.e., assuming that f can be approximated with a rate of convergence of 1/n, then f turns out to be a Lipschitz continuous function.

Intelligent Microfluidics Research on Relative Permeability Measurement and Prediction of Two-Phase Flow in Micropores

Relative permeability is a key index in resource exploitation, energy development, environmental monitoring, and other fields. However, the current determination methods of relative permeability are inefficient and invisible without considering wetting order and pore structure characteristics either. In this study, microfluidic experiments were designed for figuring out key factors impacting on the two-phase relative permeability. The optimized intelligent image recognition was established for saturation extraction. The deep learning was conducted for the prediction of two-phase permeability based on the inputs from microfluidic experiments and image recognition and optimized. Results revealed that phase saturation, wetting order, and pore topology were the key factors influencing the two-phase relative permeability, with the importance of 38.22%, 34.84%, and 26.94%, respectively. The deep learning-based relative permeability model performed well, with MSE < 0.05 and operational efficiency of 3 ms/epoch. Aiming at relative permeability model optimization, on the one hand, the dividing ratio of training set and testing set for flooding phase relative permeability prediction achieved the highest prediction accuracy at 7 : 3, while that for displaced phase was 6 : 4. On the other hand, tanh() activation function performed 40% more accurate than the sigmoid() activation function.

On one approach to the development of a simulator of the movement of an autonomous vehicle with training

The article deals with the use of simulators for controlling the movement of an autono-mous vehicle and development of a new simulator. The approach to creating a simulator of motion of a vehicle in the C# programming language is described. In the development for the implementation of simulation scenes used Unity 3D multi-platform tool is used in the devel-opment. The simulation uses direct propagation neural network that does not have a clear number of input level neurons, having only a constant output level, consisting of two neurons: the first one is responsible for acceleration, the second one is responsible for the ability of the car to turn left or to the right. Also, there is no clearly defined number of hidden levels and neurons located there. All this data in the simulation can be determined by the user. The input to the neural network values received from lasers. The lasers measure the distances to obstacles and feed the values to the input of the neural network. A sigmoidal activation function is implemented. To train the neural network an augmented learning algorithm is used, namely, a genetic algorithm applied to each vehicle, starting with the creation of each vehicle's own list of genes. In the network of each vehicle the number of genes is equal to the number of weights. For the first generation, the weights are set randomly. For the simulation, a generalized neural network with a large number of settings, with it is possible to change its structure: it is possible to change the number input level neurons that depend on the number of lasers at vehicle, their range, the height at which they detect interference, field of their visibility; you can change the number of hidden levels and the number of neurons that will be located there; control the mutation process used in the genetic algorithm; define the value of the mutation and the range of variation at which values can be varied; turn on and off the self-preservation, change vehicle speed, acceleration, set the maximum and minimum speed, edit the parameters responsible for the rotation of the car and its smoothness. Implementation of the project is provided on GitHub. The simulator can be downloaded by any developer from GitHub and can be used to implement and test various neural network training algorithms, including work of your own design.

Effect of activation function in modeling the nexus between carbon tax, CO2 emissions, and gas-fired power plant parameters

• The nexus between carbon tax, CO 2 emissions and parameters in gas-fired power plant was modeled. • The optimized networks consist of 3 input units, 15 hidden neurons, and 1 output unit. • The network performance in modeling the carbon tax prediction resulted in R 2 greater than 0.9. • The gas flow rate had the most significant effect on the predicted carbon tax. Huge emissions of carbon dioxide (CO 2 ) from the utilization of fossil fuel for power generation has significantly contributed to global warming. In view of this, technological pathways have been initiated to mitigate the effect of CO 2 emissions through capture, storage, and utilization. Besides, there is an increasing acceptance of carbon tax which is levied in the proportion of carbon emissions from the utilization of fossil fuel. In this study, the nexus between carbon tax, equivalent CO 2 emissions from the gas-fired power plant, natural gas flow rate, and air-to-fuel ratio was modeled using a perceptron neural network. The effect of various combinations of identity, hyperbolic tangent, and sigmoid activation functions at the hidden and outer layer of the neural network on the performance of the models was investigated. The various network configurations were trained using the Levenberg-Marquardt algorithm with the network errors backpropagated to enhance the performance. The optimized networks consist of three input units, 15 hidden neurons, and one output unit. The network performance in modeling the carbon tax prediction resulted in R 2 of 0.999, 0.999, 0.999, 0.998, and 0.999 for model 1, model 2, model 3, model 4, and model 5, respectively which is an indication that the calculated carbon tax was strongly correlated with the predicted values. The prediction errors of 0.019, 0.009, 0.002, 0.016, 0.002 obtained from model 1, model 2, model 3, model 4, and model 5, respectively revealed the robustness of the models in predicting the carbon tax with minimum error. Among the various configurations investigated, the perceptron neural network configured with hyperbolic tangent and sigmoid activation function at the hidden and outer layers, as well as the configuration with sigmoid activation functions at the hidden and outer layers, offer the best performance. The sensitivity analysis shows that the flow rate of the natural gas had the most significant effect on the predicted carbon tax.

Denigrate Comment Detection in Low-Resource Hindi Language Using Attention-Based Residual Networks

Cyberspace has been recognized as a conducive environment for use of various hostile, direct, and indirect behavioural tactics to target individuals or groups. Denigration is one of the most frequently used cyberbullying ploys to actively damage, humiliate, and disparage the online reputation of target by sending, posting, or publishing cruel rumours, gossip, and untrue statements. Previous pertinent studies report detecting profane, vulgar, and offensive words primarily in the English language. This research puts forward a model to detect online denigration bullying in low-resource Hindi language using attention residual networks. The proposed model Hindi Denigrate Comment–Attention Residual Network (HDC-ARN) intends to uncover defamatory posts (denigrate comments) written in Hindi language which stake and vilify a person or an entity in public. Data with 942 denigrate comments and 1499 non-denigrate comments is scraped using certain hashtags from two recent trending events in India: Tablighi Jamaat spiked Covid-19 (April 2020, Event 1) and Sushant Singh Rajput Death (June 2020: Event 2). Only text-based features, that is, the actual content of the post, are considered. The pre-trained word embedding for Hindi language from fastText is used. The model has three ResNet blocks with an attention layer that generates a post vector for a single input, which is passed through a sigmoid activation function to get the final output as either denigrate (positive class) or non-denigrate (negative class). An F-1 score of 0.642 is achieved on the dataset.