Two-step Neural Network Research Articles

Electromagnetic Imaging of Uniaxial Objects by Two-Step Neural Network

The integration of electromagnetic imaging technology with the Internet of Things plays an important role in fields as diverse as healthcare, geophysics, and industrial diagnostics. This paper presents a novel two-step neural network architecture to solve the electromagnetic imaging for uniaxial objects which can be used in the Internet of Things. We incident TM and TE waves to unknown objects and receive the scattered fields. In order to reduce the training difficulty, we first input the gathered scattered field information into a deep convolutional neural network (DCNN) to obtain the preliminary guess. In the second step, we feed the guessed image into the convolutional neural network (CNN) to reconstruct high-resolution images. Our numerical results demonstrate the real-time imaging capability of our proposed two-step method in reconstructing high-contrast scatterers.

Exploring a Novel Mexican Sign Language Lexicon Video Dataset

This research explores a novel Mexican Sign Language (MSL) lexicon video dataset containing the dynamic gestures most frequently used in MSL. Each gesture consists of a set of different versions of videos under uncontrolled conditions. The MX-ITESO-100 dataset is composed of a lexicon of 100 gestures and 5000 videos from three participants with different grammatical elements. Additionally, the dataset is evaluated in a two-step neural network model as having an accuracy greater than 99% and thus serves as a benchmark for future training of machine learning models in computer vision systems. Finally, this research provides an inclusive environment within society and organizations, in particular for people with hearing impairments.

1024-ary composite OAM shift keying for free-space optical communication system decoded by a two-step neural network.

The demand for high-dimensional encoding techniques for communication systems is increasing. Vortex beams carrying orbital angular momentum (OAM) provide new degrees of freedom for optical communication. In this study, we propose an approach for increasing the channel capacity of free-space optical communication systems by integrating superimposed orbital angular momentum (OAM) states and deep learning techniques. We generate composite vortex beams with topological charges ranging from -4 to 8 and radial coefficients ranging from 0 to 3. A phase difference among each OAM state is introduced to significantly increase the number of available superimposed states, achieving up to 1024-ary codes with distinct features. To accurately decode the high-dimensional codes, we propose a two-step convolutional neural network (CNN). The first step is to make a coarse classification of the codes, while the second step is to finely identify the code and achieve decoding. Our proposed method demonstrates 100% accuracy achieved for the coarse classification after 7 epochs, 100% accuracy achieved for the fine identification after 12 epochs, and 99.84% accuracy achieved for testing, which is much faster and more accurate than one-step decoding. To demonstrate the feasibility of our method, we successfully transmitted a 24-bit true-color Peppers image once with a resolution of 64 × 64 in the laboratory, yielding a bit error rate of 0.

Two-Step Neural-Network-Based Fault Isolation for Stochastic Systems

This paper studies a fault isolation method for an optical fiber vibration source detection and early warning system. We regard the vibration sources in the system as faults and then detect and isolate the faults of the system based on a two-step neural network. Firstly, the square root B-spline expansion method is used to approximate the output probability density functions. Secondly, the nonlinear weight dynamic model is established through a dynamic neural network. Thirdly, the nonlinear filter and residual generator are constructed to estimate the weight, analyze the residual, and estimate the threshold, so as to detect, diagnose, and isolate the faults. The feasibility criterion of fault detection and isolation is given by using some linear matrix inequalities, and the stability of the estimation error system is proven according to the Lyapunov theorem. Finally, simulation experiments based on a optical fiber vibration source system are given to verify the effectiveness of this method.

Distance Measurement by Neural Network Learning of Near-Field Microwave Reflection Spectra

Microwave-based distance measurements are limited depending on the sensing environment, such as the propagation medium and surrounding obstacles, and the complex environment also affects the measurement performance. To tackle this problem, we propose a method for predicting the distance based on the artificial neural network learning of near-field microwave reflection spectra. In principle, the spectral data is expected to contain a signature of the distance of the target object. Based on this, we proposed a two-step neural network to extend the measurable distance range while ensuring prediction performance. The first step is to predict the coarse range of the target by classification, and the next is to predict the precise distance value through multidimensional regression within that coarse range. The method was verified through experiments to predict the position of an object in an underwater environment, which was difficult to measure with conventional methods.

SCANN: Synthesis of Compact and Accurate Neural Networks

Deep neural networks (DNNs) have become the driving force behind recent artificial intelligence (AI) research. With the help of a vast amount of training data, neural networks can perform better than traditional machine learning algorithms in many applications. An important problem with implementing a neural network is the design of its architecture. Typically, such an architecture is obtained manually by exploring its hyperparameter space and kept fixed during training. This approach is both time consuming and inefficient. Another issue is that modern neural networks often contain millions of parameters, whereas many applications require small inference models due to imposed resource constraints, such as energy constraints on battery-operated devices. However, efforts to migrate DNNs to such devices typically entail a significant loss of classification accuracy. To address these challenges, we propose a two-step neural network synthesis methodology, called DR+SCANN, that combines two complementary approaches to design compact and accurate DNNs. At the core of our framework is the SCANN methodology that uses three basic architecture-changing operations, namely, connection growth, neuron growth, and connection pruning, to synthesize feedforward architectures with arbitrary structure. These neural networks are not limited to the multilayer perceptron structure. SCANN encapsulates three synthesis methodologies that apply a repeated grow-and-prune paradigm to three architectural starting points. DR+SCANN combines the SCANN methodology with dataset dimensionality reduction to alleviate the curse of dimensionality. We demonstrate the efficacy of SCANN and DR+SCANN on various image and nonimage datasets. We evaluate SCANN on MNIST, CIFAR-10, and ImageNet benchmarks. Without any loss in accuracy, SCANN generates a <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$46.3\times $ </tex-math></inline-formula> smaller network than the LeNet-5 Caffe model. We also compare SCANN-synthesized networks with a state-of-the-art fully connected (FC) feedforward model for MNIST, and show <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$20\times $ </tex-math></inline-formula> ( <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$19.9\times $ </tex-math></inline-formula> ) reduction in the number of parameters (floating-point operations) with little drop in accuracy. For the CIFAR-10 dataset, we target AlexNet and VGG-16 baseline architectures. SCANN reduces the number of parameters in AlexNet by <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$10.1\times $ </tex-math></inline-formula> without any drop in accuracy. It reduces the number of parameters in the FC layers of VGG-16 to only 2.5k while increasing accuracy by 1.05%. On the ImageNet dataset, for the VGG-16 and MobileNetV2 architectures, we reduce network parameters by <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$8.0\times $ </tex-math></inline-formula> and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$1.3\times $ </tex-math></inline-formula> , respectively, with a similar or improved performance over their respective baselines. We also evaluate the efficacy of using dimensionality reduction alongside SCANN (DR+SCANN) on nine small-to-medium-size datasets. Using this methodology enables us to reduce the number of connections in the network by up to <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$5078.7\times $ </tex-math></inline-formula> (geometric mean: <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$82.1\times $ </tex-math></inline-formula> ), with little to no drop in accuracy. On seven out of nine datasets, we show 0.41%–10.09% accuracy improvements over the FC baseline models. We also show that our synthesis methodology yields neural networks that are much better at navigating the accuracy versus energy efficiency space. This can enable neural network-based inference even on Internet-of-Things sensors.

Soft Sensing and Multi-objective Optimization of Pulp Washing Process Based on Data-Driven Technology

Due to the multi-variable, time-delay, and nonlinear characteristics of the pulp washing process, it is difficult to accurately measure and optimize the process. An asymmetrical nonlinear control system can be decomposed into a group of low-dimensional subsystems, which brings great convenience to the analysis and design of the system. In this paper, the concept of symmetry is used to simplify nonlinear optimal control problems, and data-driven theory is used to solve optimal policy problems. This paper proposes a data-driven operating model optimization method to model and optimizes the pulp washing process. The most important quality indicators of pulp washing performance are the alkali residue in washing pulp and the Baumé degree of the extracted black liquor. Considering the difficulty of modeling, online measurement of these indicators, two-step neural network, and multiple logistic regression were used to build a prediction model for soda water and Baumé degree. The mathematical model of the washing process can be identified, and the indicators meet the production requirements. With the goal of better product quality, low cost, and low energy consumption, based on the optimized operation mode database, the ant colony optimization (ACO) algorithm was used to solve the multi-objective problem. Theoretical analysis and practical application were carried out, and the optimal control system of the pulp washing process was designed. The actual results showed that the pulp output was increased by 20%, and the water consumption was reduced by nearly 30%. This method is effective during pulp washing.

Estimating cluster masses from SDSS multiband images with transfer learning

ABSTRACTThe total masses of galaxy clusters characterize many aspects of astrophysics and the underlying cosmology. It is crucial to obtain reliable and accurate mass estimates for numerous galaxy clusters over a wide range of redshifts and mass scales. We present a transfer-learning approach to estimate cluster masses using the ugriz-band images in the SDSS Data Release 12. The target masses are derived from X-ray or SZ measurements that are only available for a small subset of the clusters. We designed a semisupervised deep learning model consisting of two convolutional neural networks. In the first network, a feature extractor is trained to classify the SDSS photometric bands. The second network takes the previously trained features as inputs to estimate their total masses. The training and testing processes in this work depend purely on real observational data. Our algorithm reaches a mean absolute error (MAE) of 0.232 dex on average and 0.214 dex for the best fold. The performance is comparable to that given by redMaPPer, 0.192 dex. We have further applied a joint integrated gradient and class activation mapping method to interpret such a two-step neural network. The performance of our algorithm is likely to improve as the size of training data set increases. This proof-of-concept experiment demonstrates the potential of deep learning in maximizing the scientific return of the current and future large cluster surveys.

Identification of Extracapsular Extension in Head and Neck Cancer Patients Using Artificial Intelligence.

<h3>Purpose/Objective(s)</h3> Routine diagnostic CT scan guides the management of head and neck squamous cell carcinoma (HNSCC)) by identifying tumor and lymph node features. Nevertheless, lymph node extracapsular extension (ECE) remains challenging to detect by clinicians even using modern imaging techniques. It is essential to detect the occurrence of ECE as it changes staging and management for the patients. Current clinical ECE detection relies on clinical examination, radiological identification, and histopathological confirmation. Therefore, we aim to perform ECE identification automatically using a gradient-based class activation mapping (Grad-CAM) technique to guide the deep learning model to focus on the regions highly suspicious of ECE. The presence or absence of ECE has been analyzed and correlated with gold standard histopathological findings. <h3>Materials/Methods</h3> This research proposes a novel deep learning framework to detect ECE from 3D head and neck CT scans. The research problem is treated as a classification task, i.e., to label patients' CT scans with ECE positive or negative categories. Grad-CAM can provide visual explanations for deep neural networks by indicating the critical regions in the image for prediction. The proposed network has two convolutional neural networks (CNNs) involved. The first CNN model acts as a preprocessing step. The Grad-CAM method is applied to the trained CNN to highlight and extract the volumes of interest (VOIs) related to the class ECE positive. The second CNN model takes the extracted VOIs as model input for ECE positive/negative classification. The data used for the second CNN training does not involve that in the first step training process. The studied dataset includes 130 patients' CT scans. Cross-validation has been applied in our experimental tests. <h3>Results</h3> The experimental results have demonstrated that our model can identify ECE positive and negative patients based on ten-fold cross-validation. We have achieved 95.03% test accuracy and 90.90% test AUC in ECE classification. Furthermore, ECE probability heatmaps are generated by Grad-CAM. <h3>Conclusion</h3> This study demonstrates that a deep learning model guided by the Grad-CAM approach can be used for ECE detection. The proposed two-step neural networks can help capture sufficient features and achieve high classification accuracy. The heatmaps can be used as a tool for radiologists to assist in the identification of ECE. The outcome of this research is expected to promote the implementation of artificial intelligence for head and neck cancer ECE identification diagnosis in the radiology computer vision field.

Machine Learning for Improving Stellar Image-based Alignment in Wide-field Telescopes

Stellar images will deteriorate dramatically when the sensitive elements of wide-field survey telescopes are misaligned during an observation, and active alignment is the key technology to maintain the high resolution of wide-field sky survey telescopes. Instead of traditional active alignment based on field-dependent wave front errors, this work proposes a machine learning alignment metrology based on stellar images of the scientific camera, which is more convenient and higher speed. We first theoretically confirm that the pattern of the point-spread function over the field is closely related to the misalignment status, and then the relationships are learned by two-step neural networks. After two-step active alignment, the position errors of misalignment parameters are less than 5 μm for decenter and less than 5″ for tip-tilt in more than 90% of the cases. The precise alignment results indicate that this metrology provides a low-cost and high-speed solution to maintain the image quality of wide-field sky survey telescopes during observation, thus implying important significance and broad application prospects.

Machine learning-based aerosol characterization using OCO-2 O2 A-band observations

Aerosol scattering influences the retrieval of the column-averaged dry-air mole fraction of CO2 (XCO2) from the Orbiting Carbon Observatory-2 (OCO-2). This is especially true for surfaces with reflectance close to a critical value where there is very low sensitivity to aerosol loading. A spectral sorting approach was introduced to improve the characterization of aerosols over coastal regions. Here, we generalize this procedure to land surfaces and use a two-step neural network to retrieve aerosol parameters from OCO-2 measurements. We show that, by using a combination of radiance measurements in the continuum and inside the absorption band, both the aerosol optical depth and layer height, as well as their uncertainties, can be accurately predicted. Using the improved aerosol estimates as a priori, we demonstrate that the accuracy of the XCO2 retrieval can be significantly improved compared to the OCO-2 Level-2 Standard product. Furthermore, using simulated observations, we obtain estimates of the error in the retrieved XCO2. These simulations indicate that the bias-corrected OCO-2 Lite data, which is used for flux inversions, may have remaining biases due to interference of aerosol effects.

Abstract 10961: Integrated Coronary Vascular Disease Diagnosis System Based on Two-Step Neural Network Algorithm Trained by Synthetic Model and Patient Biometric Features

Introduction: Fractional flow reserve (FFR) is the golden standard that determines whether coronary artery disease (CAD) surgery is necessary or not. Invasive FFR uses pressure wire to obtain the pressure ratio of proximal and distal lesion. However, the burden from high cost and time due to invasive process makes it difficult to perform diagnosis. To overcome these limitations, FFR can be measured non-invasive through CFD analysis based on geometry models obtained from observational equipment such as coronary computer tomography (CT). However, obtaining this FFR value consumes considerable time and computer resources in the process of vessel shape extraction and CFD. In this study, we propose a non-invasive artificial intelligence-based FFR prediction technology. Methods &amp; Results: The algorithms used in this paper were separated into two steps. The first algorithm is to compute flow features from geometric features trained by synthetic models and CFD analysis. In the second step, we trained classification algorithms and regression algorithms using the flow properties obtained from the first step and patient biometric information. Classification algorithms are used to determine whether to invasive operate, and regression algorithms are used to predict FFR values. The algorithms used are 6 machine learning and 2 neural networks. To improve the performance of the algorithms, outlier data were removed through the IQR method. Among the classification algorithms, random forest had the best accuracy of 80.1%, and in the regression algorithm, artistic neural network had the lowest mean absolute error of 0.04. Furthermore, we analyze the algorithm performance within the gray zone to prove that the algorithm can provide guidance in the case of uncertain situations. Conclusions: Algorithms learned in a two-step manner determined whether the lesion need a surgery or not and predicted the FFR value. In particular, we presented guidelines for building FFR prediction algorithms with good prediction performance.

Bridge Damage Detection in Presence of Varying Temperature Using Two-Step Neural Network Approach

Abstract The dynamic properties of bridges can be affected not only through damage but also from ambient uncertainty. False-positive or negative alarms may be raised if environmental effects are no...

A Two-stage Deep Learning Detection Classifier for the ATLAS Asteroid Survey

In this paper we present a two-step neural network model to separate detections of solar system objects from optical and electronic artifacts in data obtained with the “Asteroid Terrestrial-impact Last Alert System” (ATLAS), a near-Earth asteroid sky survey system. A convolutional neural network is used to classify small “postage-stamp” images of candidate detections of astronomical sources into eight classes, followed by a multi-layered perceptron that provides a probability that a temporal sequence of four candidate detections represents a real astronomical source. The goal of this work is to reduce the time delay between Near-Earth Object (NEO) detections and submission to the Minor Planet Center. Due to the rare and hazardous nature of NEOs, a low false negative rate is a priority for the model. We show that the model reaches 99.6% accuracy on real asteroids in ATLAS data with a 0.4% false negative rate. Deployment of this model on ATLAS has reduced the amount of NEO candidates that astronomers must screen by 90%, thereby bringing ATLAS one step closer to full autonomy.

High Accuracy Modeling for Solar PV Power Generation Using Noble BD-LSTM-Based Neural Networks with EMA

More accurate self-forecasting not only provides a better-integrated solution for electricity grids but also reduces the cost of operation of the entire power system. To predict solar photovoltaic (PV) power generation (SPVG) for a specific hour, this paper proposes the combination of a two-step neural network bi directional long short-term memory (BD-LSTM) model with an artificial neural network (ANN) model using exponential moving average (EMA) preprocessing. In this study, four types of historical input data are used: hourly PV generation for one week (168 h) ahead, hourly horizontal radiation, hourly ambient temperature, and hourly device (surface) temperature, downloaded from the Korea Open Data Portal. The first strategy is employed using the LSTM prediction model, which forecasts the SPVG of the desired time through the data from the previous week, which is preprocessed to smooth the dynamic SPVG using the EMA approach. The SPVG was predicted using the LSTM model according to the trend of the previous time-series data. However, slight errors still occur because the weather condition of the time is not reflected at the desired time. Therefore, we proposed a second strategy of an ANN model for more accurate estimation to compensate for this slight error using the four inputs predicted by the LSTM model. As a result, the LSTM prediction model with the ANN estimation model using EMA preprocessing exhibited higher accuracy in performance than other options for SPVG.

Data-Driven Optimal Control for Pulp Washing Process Based on Neural Network

Pulp washing process has the features of multivariate, time delay, nonlinearity. Considering the difficulties of modeling and optimal control in pulp washing process, a data-driven operational-pattern optimization method is proposed to model and optimize the pulp washing process in this paper. The most important quality indexes of pulp washing performance are residual soda in the washed pulp and Baume degree of extracted black liquor. Considering the difficulties of modeling, online measurement of these indexes, two-step neural networks, and multivariate logistic regression are used to establish the prediction models of residual soda and Baume degree. The mathematical model of the washing process can be identified, and the indexes can meet the production requirements. In the target of better product quality, low cost, and low energy consumption, a multiobjective problems is solved by ant colony optimization algorithm based on the optimized operational-pattern database. It shows that the theoretical analyses are correct and the practical applications are feasible, optimization control system has been designed for the pulp washing process, and the practical results show that pulp production increased by 20% and water consumption decreased by nearly 30%. This method is effective in the pulp washing process.

QoE Evaluation Methods for 360-Degree VR Video Transmission

Virtual reality (VR) videos, especially 360-degree VR videos, have attracted substantial research interests with the explosion of mobile VR devices in recent years. In this work, we propose a VR quality of experience (QoE) evaluation framework including online, offline and mixed scenarios that can meet the requirement of real applications. Under this framework, we further develop a subjective evaluation method and an objective QoE evaluation model based on network transmission parameters. A VR panoramic video database is established with nine different contents. This database contains 468 different samples coming from nine original contents. Then, in the subjective evaluation part, we propose to adopt four dimensions, i.e., quality score, immersion score, non-spinning sensation score and global score for the subjective evaluation purpose. These dimensions can meet the interaction and immersion features of the VR video service. In the objective evaluation part, an improved two-step neural network (INN) model is developed by using the features of the physiological psychology and cognitive neurology. This model captures the intrinsic link between the initial input network parameters and the final perception in VR transmission. Our results suggest that the proposed INN can achieve a higher pearson correlation coefficients than the other methods. Especially, it increases by 8.9% than the traditional NN method. Our work can make up for the shortage of traditional evaluation methods in term of the VR QoE subjective and objective evaluation in transmission scopes.

A Machine Learning Approach to Achieving Energy Efficiency in Relay-Assisted LTE-A Downlink System.

In recent years, Energy Efficiency (EE) has become a critical design metric for cellular systems. In order to achieve EE, a fine balance between throughput and fairness must also be ensured. To this end, in this paper we have presented various resource block (RB) allocation schemes in relay-assisted Long Term Evolution-Advanced (LTE-A) networks. Driven by equal power and Bisection-based Power Allocation (BOPA) algorithm, the Maximum Throughput (MT) and an alternating MT and proportional fairness (PF)-based SAMM (abbreviated with Authors’ names) RB allocation scheme is presented for a single relay. In the case of multiple relays, the dependency of RB and power allocation on relay deployment and users’ association is first addressed through a k-mean clustering approach. Secondly, to reduce the computational cost of RB and power allocation, a two-step neural network (NN) process (SAMM NN) is presented that uses SAMM-based unsupervised learning for RB allocation and BOPA-based supervised learning for power allocation. The results for all the schemes are compared in terms of EE and user throughput. For a single relay, SAMM BOPA offers the best EE, whereas SAMM equal power provides the best fairness. In the case of multiple relays, the results indicate SAMM NN achieves better EE compared to SAMM equal power and BOPA, and it also achieves better throughput fairness compared to MT equal power and MT BOPA.

Accelerating and stabilizing the vapor-liquid equilibrium (VLE) calculation in compositional simulation of unconventional reservoirs using deep learning based flash calculation

The flash calculation with large capillary pressure often turns out to be time-consuming and unstable. Consequently, the compositional simulation of unconventional oil/gas reservoirs, where large capillary pressure exists on the vapor-liquid phase interface due to the narrow pore channel, becomes a challenge to traditional reservoir simulation techniques. In this work, we try to resolve this issue by combining deep learning technology with reservoir simulation. We have developed a compositional simulator that is accelerated and stabilized by stochastically-trained proxy flash calculation.We first randomly generated 300,000 data samples from a standalone physical flash calculator. We have constructed a two-step neural network, in which the first step is the classify the phase condition of the system and the second step is to predict the concentration distribution among the determined phases. Each network consists of four hidden layers in between the input layer and the output layer. The network is trained by Stochastic Gradient Descent (SGD) method with 100 epochs.With given temperature, pressure, feed concentration pore radius, the trained network predicts the phases and concentration distribution in the system with very low computational cost. Our results show that the accuracy of the network is above 97% in the metric of mean absolute percentage error. The predicted result is used as the initial guess of the flash calculation module in the reservoir simulator. With the implementation of the deep learning based flash calculation module, the speed of the simulator has been effectively increased and the stability (in the manner of the ratio of convergence) has been improved as well.

Multi-Axial Force/Torque Sensor Calibration Method Based on Deep-Learning

The multi-axial force/torque sensor for advanced robotic applications utilizes the internal strain of deformable structure for its measurement. Thus, there exist coupling and errors due to nonlinearity caused from the variation of the electric signal with respect to the strain. This paper presents a method of calibrating six-axis force/torque sensors with high accuracy based on deep learning. The method can solve the aforementioned problems easily by using deep-neural network (DNN). After performing the structural analysis of the sensor, generalized equations of the electric signal is derived, which leads it to the basic DNN structure, and optimization is performed. Training and test data are prepared by using a dummy and a reference sensor. Then, the proposed method is validated by comparing the three results obtained from the linear transformation method based on least-square method, two-step neural-network, and the DNN-based method for each untrained test data set.