Optical Memory and Neural Networks最新文献

This paper presents a theoretical work of a new device concept for frequency division demultiplexing with excellent performance based on waveguides system containing segments and loops in the presence of two geometrics defects. This system permits the separation of two frequency, based on 1D photonic waveguides loops structures. The system under consideration possesses a Y‑shaped demultiplexer configuration, consisting of a single input and two output channels (transmission lines). Each output channel contains an alternating unit cell consisting of a segment and a loop. The creation of a geometrical defect at the segment level in the middle of each output line allows the creation of two defect modes inside the bandgaps. The numerical results show that this demultiplexer system is able to separate two signals (electromagnetic waves) of different frequencies and guide each signal through an output channel. We perform the analytical calculation of the transmission rates T₁, T₂, and reflection R using the interface response theory, which is based on Green’s function method for the proposed demultiplexer system. The proposed device offers high transmission efficiency, high quality factor and a large frequency difference between defect modes, hence, it is highly desirable for frequency division demultiplexing applications.

The paper investigates algorithms using long intensity gradients for georeferencing of Earth remote sensing data. The case is considered in which one “reliable” referenced set of remote sensing data is already known for a particular area. New input data are referenced to this “reliable” set by detecting resemblant fragments in the “relible” data set and new remote sensing data. A set of pairs of resemblant fragments makes it possible to calculate the transformation parameters of new data. To increase the efficiency of resemblant fragments detection, we go to the space of long intensity gradients, which makes the georeferencing method more stable to admissible differences between resemblant fragments. The paper considers a few algorithms of going to the long gradient space and compares them. The computaional experiment provides grounds for recommending the best way of going to the long gradient space.

The 1/t Wang-Landau algorithm is analyzed from the viewpoint of execution time and accuracy when it is used in computations of the density of states of a two-dimensional Ising model. We find that the simulation results have a systematic error, the magnitude of which decreases with increasing the lattice size. The relative error has two maxima: the first one is located near the energy of the ground state, and the second maximum corresponds to the value of the internal energy at the critical point. We demonstrate that it is impossible to estimate the execution time of the 1/t Wang-Landau algorithm in advance when simulating large lattices. The reason is that when the final value of the modification factor was reached, the criterion for transition to mode 1/t was not met. The simultaneous calculations of the density of states for energy and magnetization are shown to lead to higher accuracy in estimating statistical moments of internal energy.

Recently, there has been an increased interest in NeRF methods which reconstruct differentiable representation of three-dimensional scenes. One of the main limitations of such methods is their inability to assess the confidence of the model in its predictions. In this paper, we propose a new neural network model for the formation of extended vector representations, called uSF, which allows the model to predict not only color and semantic label of each point, but also estimate the corresponding values of uncertainty. We show that with a small number of images available for training, a model that quantifies uncertainty performs better than a model without such functionality. Code of the uSF approach is publicly available at https://github.com/sevashasla/usf/.

Analyzing hyperspectral images poses a non-trivial challenge due to various challenges. To overcome most of these challenges one of the widely employed approach involves utilizing indices, such as the Normalized Difference Vegetation Index (NDVI). Indices provide a powerful means to distill complex spectral information into meaningful metrics, facilitating the interpretation of specific features within the hyperspectral domain. Moreover, the indices are usually easy to compute. However, creating indices for discerning arbitrary data classes within an image proves to be a challenging task. In this paper, we present an algorithm designed to automatically generate lightweight descriptors, suited for discerning between arbitrary classes in hyperspectral images. These lightweight descriptors within the algorithm are characterized by indices derived from selected informative layers. Our proposed algorithm streamlines the descriptor generation process through a multi-step approach. Firstly, it employs Principal Component Analysis (PCA) to transform the hyperspectral image into a three-channel representation. This transformed image serves as input for a Segment Anything Model (SAM). The neural network outputs a labeled map, delineating different classes within the hyperspectral image. Subsequently, our Informative Index Formation algorithm (INDI) utilizes this labeled map to systematically generate a set of lightweight descriptors. Each descriptor within the set is adept at distinguishing a specific class from the remaining classes in the hyperspectral image. The paper demonstrates the practical application of the developed algorithm for hyperspectral image segmentation.

Reducing the amount of car accidents and the deaths that result from them requires close monitoring of drivers’ health and alertness. Identifying driver weariness has been a major practical concern and problem in recent years. A number of machine learning algorithms have been used for monitoring the driver’s health system, even though accurate and early identification is more challenging. In order to overcome this issues, vehicle driver health is monitored using wearable ECG based on an optimized Deep Belief Network (DBN) is proposed. The collected ECG raw signal is pre-processed using a notch filter and high pass filter and an adaptive sliding window to improve the signal quality. After that, Wavelet Packet Decomposition (WPD) and the Short Time Fourier Transform (SIFT) are used to extract features from the pre-processed signal. It enables for the extraction of both time and frequency domain data. In order to classify whether a driver is fit to drive, is under stress, or has a heart condition, the extracted statistical features are sent for further classification using an optimized Deep Belief Neural Network (DBN). The walrus optimization technique is utilized to set the learning rate of the DBN classifier in an optimal manner. To prevent collisions between vehicles, the driver will be alerted via a buzzer system in the event of stress or heart problems. According to the results of the experimental research, the proposed technique achieves 95.1% accuracy, 92.5% precision, 96.5% specificity, 93% of recall, and 92.7% of the f1-score. Thus, the driver health monitoring system can be accurately detected using this automated model.

Plant diseases can harm crops and reduce the amount of food that can be cultivated, which is problematic for farmers. Technology is being utilized to develop computer-based programs that can recognize plant diseases and assist farmers in making better decisions after identifying plant leaf diseases. In most of these models, machine learning algorithms are applied, to make predictions about potential plant diseases using mathematical models and neural networks. Many researchers discussed the variants of DNN and CNN algorithms to solve the discussed problems and gave better results. In this paper, the novel approach is discussed and implemented where the plant disease is identified whether the plant leaf captured image has a noisy background or not; or whether the leaf image is segmented or not. The authors developed an adaptive algorithm which gives the results in two phases: the classification of the plant disease based on the original input leaf image and secondly, the identification of plant leaf disease after applying the segmentation process. The result of this two-phase proposed model is analyzed and compared with existing popular models like AlexNet, ResNet-50, and the EffNet the results are convincing. The proposed model has 97.39% accuracy when the noiseless image is taken; while the 90.26% accuracy is there, in case of noisy background image as an input; and the results are outstanding, if the authors are applying their segmentation-based AH-CNN model on the noisy real-time image, the accuracy is 95.27%.

Lung cancer is the most common cancer and the primary reason for cancer related fatalities globally. Lung cancer patients have a 14% overall survival rate. If the cancer is found in the early stages, the lives of patients with the disease may be preserved. A variety of conventional machine and deep learning algorithms have been developed for the effective automatic diagnosis of lung cancer. But they still have issues with recognition accuracy and take longer to analyze. To overcome these issues, this paper presents deep learning assisted Squeeze and Excitation Convolutional Neural Networks (SENET) to predict lung cancer on computed tomography images. This paper uses lung CT images for prediction. These raw images are preprocessed using Adaptive Bilateral Filter (ABF) and Reformed Histogram Equalization (RHE) to remove noise and enhance an image’s clarity. To determine the tunable parameters of the RHE approach Tuna Swam optimization algorithm is used in this proposed method. This preprocessed image is then given to the segmentation process to divide the image. This proposed approach uses the Chan vese segmentation model to segment the image. Segmentation output is then fed into the classifier for final classification. SENET classifier is utilized in this proposed approach to final lung cancer prediction. The outcomes of the test assessment demonstrated that the proposed model could identify lung cancer with 99.2% accuracy, 99.1% precision, and 0.8% error. The proposed SENET system predicts CT scanning images of lung cancer successfully.

The paper introduces two novel activation functions known as modExp and modExp_m. The activation functions possess several desirable properties, such as being continuously differentiable, bounded, smooth, and non-monotonic. Our studies have shown that modExp and modExp_m consistently outperform ReLU and other activation functions across a range of challenging datasets and complex models. Initially, the experiments involve training and classifying using a multi-layer perceptron (MLP) on benchmark data sets like the Diagnostic Wisconsin Breast Cancer and Iris Flower datasets. Both modExp and modExp_m demonstrate impressive performance, with modExp achieving 94.15 and 95.56% and modExp_m achieving 94.15 and 95.56% respectively, when compared to ReLU, ELU, Tanh, Mish, Softsign, Leaky ReLU, and TanhExp. In addition, a series of experiments were carried out on five different depths of deeper neural networks, ranging from five to eight layers, using MNIST datasets. The modExp_m activation function demonstrated superior performance accuracy on various neural network configurations, achieving 95.56, 95.43, 94.72, 95.14, and 95.61% on wider 5 layers, slimmer 5 layers, 6 layers, 7 layers, and 8 layers respectively. The modExp activation function also performed well, achieving the second highest accuracy of 95.42, 94.33, 94.76, 95.06, and 95.37% on the same network configurations, outperforming ReLU, ELU, Tanh, Mish, Softsign, Leaky ReLU, and TanhExp. The results of the statistical feature measures show that both activation functions have the highest mean accuracy, the lowest standard deviation, the lowest Root Mean squared Error, the lowest variance, and the lowest Mean squared Error. According to the experiment, both functions converge more quickly than ReLU, which is a significant advantage in Neural network learning.

The paper studies the properties of a fully connected neural network built around phase neurons. The signals traveling through the interconnections of the network are unit pulses with fixed phases. The phases encoding the components of associative memory vectors are distributed at random within the interval [0, 2π]. The simplest case in which the connection matrix is defined according to Hebbian learning rule is considered. The Chernov–Chebyshev technique, which is independent of the type of distribution of encoding phases, is used to evaluate the recognition error. The associative memory of this type of network is shown to be four times as large as that of a conventional Hopfield-type network using binary patterns. Correspondingly, the radius of the domain of attraction is also four times larger.