Optical Memory and Neural Networks最新文献

Social media rumours significantly challenge societal discourse, demanding effective detection mechanisms. Existing automated rumour detection methods primarily rely on topological data, yet computational complexity and managing large datasets remain formidable obstacles. This study proposes a novel neural network approach utilising graphical structures to address these challenges and enhance rumour detection efficiency. This study suggests a novel neural network approach to improve rumour detection efficiency using graphical structures from the PHEME dataset. The strategy aims to improve classifier performance by transforming tweeting graphs into distinct binary trees, enabling the learning of structural information’s propagation and dispersion. This makes it possible to build meta-tree paths that record and capture local structural information. The model learns global structural representations using BERT on these pathways. The approach also incorporates user relationships and content associations utilizing a bidirectional graph convolutional network encoder to learn node-level representations. The final node-level representation is synthesised by combining user and content embeddings. A fusion approach combines the structural and node-level representations, passing through a fully connected layer and a Softmax layer for rumour detection. This proposed model outperforms the existing models, with an accuracy of over 93% without cross-validation and more than 95% with cross-validation. Experimental validation demonstrates the effectiveness of the suggested approach in rumour detection over social media, offering a promising solution to mitigate the impact of misinformation and rumours in online discourse.

This paper addresses the problem of object keypoint detection from a single image using modern machine learning methods. Keypoint detection has been extensively studied for human pose estimation, and thus, the study compares deep convolutional neural networks that effectively solve this task. Given the challenge of adapting methods to different object types, special attention is paid to automating the preparation of training data. A novel approach is presented, which includes generating datasets based on 3D models, automatically annotating keypoints, and capturing images of objects from various angles, scales, backgrounds, and lighting conditions. The study investigates which modern deep neural networks are the most effective for keypoint detection and explores the applicability of models trained on synthetic data to real-world scenarios.

The paper investigates an iterative approximation-based method of compressing discrete multispectral images. Less thinned multispectral images are used to approximate more thinned ones, the degree of thinning decreasing in an iterrative fashion. When a set of thinned multispectral images is used, the data redundency is eliminated by using nonredundant nested covers consisted of specially reduced thinned images. Approximation errors are rounded and stored. The paper considers an algorithm of detection and effective representation of degenerate subsets of rounded iterative approximation errors. The algotithm allows more efficient representation of rounded error subsets and higher data compression ratios. The computational experiment confirms a considerable increase in efficiency of the iterative approximation-based method of discrete multispectral data compression.

We discuss two source vector fields of Poincare-beam type that can be looked upon as optical skyrmions, i.e. topological quasiparticles. We derive explicit analytical relationships that describe projections of a three-dimensional (3D) skyrmion vector field in the source plane and skyrmion numbers, which are shown to be pro-portional to the topological charges of constituent optical vortices of the Poincare beams. We also propose a new constructive formula as an effective tool for calculating the skyrmion number via normalized Stokes vector projections rather than skyrmion vector field projections. The skyrmion numbers calculated using the familiar and newly proposed formulae coincide. Numbers of each projection of a 3D skyrmion vector field are shown to comprise a third of the full skyrmion number. The theoretical conclusions are validated by a numerical simulation. The non-uniform linear polarization in the skyrmion cross-section depends on the azimuthal angle and can be used to form a spiral microrelief due to the mass transfer of molecules on the surface of the material.

We analyze a Hopfield neural network with a quasi-diagonal connection matrix. We use the term “quasi-diagonal matrix” to denote a matrix with all elements equal zero except the elements on the first super- and sub-diagonals of the principle diagonal. The nonzero elements are arbitrary real numbers. Such matrix generalizes the well-known connection matrix of the one dimensional Ising model with open boundary conditions where all nonzero elements equal ( + 1). We present a simple description of the fixed points of the Hopfield neural network and their dependence on the matrix elements. The obtained results also allow us to analyze the cases of a) the nonzero elements constitute arbitrary super- and sub-diagonals and b) periodic boundary conditions.

Accurate vegetable yield prediction and precise fertilizer recommendations are crucial for maximizing agricultural productivity and sustainability. Advanced methodologies, including machine learning algorithms and precision agriculture tools, offer significant improvements in forecasting crop yields and optimizing nutrient application. However, traditional models often depend on extensive, high-quality datasets, which may be challenging to obtain in less-developed regions. Moreover, traditional fertilizer recommendation systems may not sufficiently adapt to real-time changes in soil conditions or crop requirements, leading to less precise nutrient management. In order to address the aforementioned problems, vegetable yield prediction and fertilizer recommendations are made using optimal machine learning and hybrid deep learning models. In this paper, the developed model collects agricultural data from a standard source. Subsequently, the collected data undergoes three pre-processing techniques to improve crop yield prediction. Data cleaning involves identifying missing or incomplete values, while data normalization ensures all features contribute equally to model training using weighted k-means and Neighbourhood averaging addresses outliers. After that pre-processed data is used for feature selection, using Relief Feature Ranking with Recursive Feature Elimination. The selected data is used for crop yield prediction and fertilizer recommendation. Physics-informed neural networks (PINN) based fruit fly optimization (IFO) algorithm is employed for predicting the yield of various vegetables like chickpeas, kidney beans, blackgram, lentil, etc. A hybrid Independent Shearlet-based Deep Belief Network (IS-DBN) is used for fertilizer recommendation. The performance metrics for vegetable prediction and fertilizer recommendation attained for the proposed model are 99.17 and 96.99% of accuracy, 91.67 and 89.65% of precision. The proposed model’s obtained values are better than those of the existing methods. Thus, the proposed optimized machine learning and hybrid deep learning approach effectively predict crop yield and fertilizer recommendation with higher accuracy.

A method for compressing multidimensional discrete arrays is studied. Based on the reduction of the extremum of post-extrapolation residuals, the method can be used to handle video and image arrays. The compression algorithm calculates post-extrapolation residuals for all samples of an initial multidimensional discrete array and then roughens these residuals to reduce the required capacity of a storage system and raise the data transmission rate. The method allows more reliable control of discrepancy between the original multidimensional discrete array and its unpacked version by reducing the extremum of roughened post-extrapolation residuals. Computer simulations validate that the use of the reduction of the extremum of post-extrapolation residuals increases the efficiency of the multidimensional data compression algorithm. The experiments also demonstrate the approach to be more efficient than other popular algorithms used for the compression of multidimensional discrete-data arrays.

Ad-hoc network is a type of wireless network, but it differs from other wireless networks in that it lacks infrastructure such as access points, routers, and other devices. While a node can communicate with every other node in the same cell in infrastructure networks, routing and the limitations of wireless communication are the main issues in ad hoc networks. But those clarifications are gave not accurate results. In order to overcome these issues, proposed traffic congestion prevention for IoT based traffic management in ad-hoc network using deep learning. This proposed method has five phases like data collection, preprocessing, feature selection, classification and decision making. The input data gathered from IoT devices in the ad-hoc network. After that, IoT features were preprocessed using missing values replacement and SMOTE resampling. Then preprocessed IoT data features to be selected using chi-square, which is used to select optimal features to avoid overfitting problems. Following that, the selected IoT features were classified with the help of deep LSTM, which is used to know whether the network is traffic or not. If the network have traffic, the data transmission is done through the traffic less path. Otherwise, the IoT data should be transmitted easily. The proposed model was designed and the performance was validated by using MATLAB software. Deep learning (DL) performance parameters such as accuracy, precision, recall, and error have values of 98.32, 98.325, 97.87, and 1.9%, respectively. Moreover, this proposed model is effective for detecting traffic congestion and which is used to prevent traffic through an ad-hoc network’s IoT based traffic management system.

The world economy, human well-being, and the health of plants and animals have all suffered greatly as a result of rising air pollution. This survey investigates four different aspects of air pollution prediction using machine learning (ML). It examines the relationship between industrial processes and emissions, concentrating on factors and industries. Predictive models that can foretell pollution levels from industrial activity are created using machine learning techniques. ML models are used to forecast the amounts of pollution associated with vehicle traffic, as automobiles play a significant role in the degradation of urban air quality. The use of ML based approaches to predict pollution levels from natural phenomena like storms of dust, lava flows, and wildfires helps preventive measures and disaster preparedness. Lastly, ML algorithms are used to anticipate pollutant emissions from a range of combustion sources, including power plants, residential heating systems, and industrial boilers. In addition to discussing the consequences for pollution management strategies, the study assesses how well machine learning algorithms predict emissions. The objective is to further advance the creation of forecasting abilities that are essential for lowering the detrimental effects of air pollution on the environment and public health by providing insights into the quickly evolving field of air pollution forecasts through ML approaches.

Recognition of Human Face expression is the most significant and challenging societal interaction tasks. Humans often convey their feelings and intentions through their facial expressions in a natural and honest manner, nonverbal communication is mostly characterized by facial expressions. Various approaches for classifying emotions and facial recognition have been established to enhance the accuracy of face recognition in the infrared images. Significant issues of recent deep FER systems include overfitting due to insufficient training data as well as expression-unrelated variables such as identification bias, head posture, and illumination. To address these challenges, the proposed model implemented a method for detecting facial expression using HAPNet segmentation and hybrid VGG16 with SVM classifier. At first, pre-processed the images using an optimized Difference of Gaussians (DOG) filter for enhancing the edges of the image and the Artificial Gorilla Troops Optimization Algorithm (GTO) is used to select the kernel size based on the maximum PSNR. Segmentation is the next step for segmenting the face using the Hybrid, Asymmetric, and Progressive Network (HAPNet) method. Landmark is detected based on Multi-Task Cascaded Convolutional Networks (MTCNN) for identifying the location of the mouth eyes, and nose. The last step is to categorize the seven emotions which are happy, sad, disgusted, surprised, angry, fearful, and neutral on faces using the hybrid VGG16 with Support Vector Machine (SVM) algorithm. The effectiveness of the proposed methodology is evaluated using the metrics of accuracy is 96.6%, positive predictive value is 93.08%, hit rate of 95.2%, selectivity of 92.5%, negative predictive value of 95.8%, and f1-score of 94.49%. Experiments on the database illustrates that the proposed approach performs better than conventional techniques for accurately identifies the expressions on the face in the thermal images.