Multiple Neural Networks Research Articles

Unmanned Aerial Vehicle Pitch Control under Delay Using Deep Reinforcement Learning with Continuous Action in Wind Tunnel Test

Nonlinear flight controllers for fixed-wing unmanned aerial vehicles (UAVs) can potentially be developed using deep reinforcement learning. However, there is often a reality gap between the simulation models used to train these controllers and the real world. This study experimentally investigated the application of deep reinforcement learning to the pitch control of a UAV in wind tunnel tests, with a particular focus of investigating the effect of time delays on flight controller performance. Multiple neural networks were trained in simulation with different assumed time delays and then wind tunnel tested. The neural networks trained with shorter delays tended to be susceptible to delay in the real tests and produce fluctuating behaviour. The neural networks trained with longer delays behaved more conservatively and did not produce oscillations but suffered steady state errors under some conditions due to unmodeled frictional effects. These results highlight the importance of performing physical experiments to validate controller performance and how the training approach used with reinforcement learning needs to be robust to reality gaps between simulation and the real world.

Research on Network Layer Recursive Reduction Model Compression for Image Recognition

ResNet has been widely used in the field of machine learning since it was proposed. This network model is successful in extracting features from input data by superimposing multiple layers of neural networks and thus achieves high accuracy in many applications. However, the superposition of multilayer neural networks increases their computational cost. For this reason, we propose a network model compression technique that removes multiple neural network layers from ResNet without decreasing the accuracy rate. The key idea is to provide a priority term to identify the importance of each neural network layer, and then select the unimportant layers to be removed during the training process based on the priority of the neural network layers. In addition, this paper also retrains the network model to avoid the accuracy degradation caused by the deletion of network layers. Experiments demonstrate that the network size can be reduced by 24.00%–42.86% of the number of layers without reducing the classification accuracy when classification is performed on CIFAR-10/100 and ImageNet.

Surrogate Safety Measures Prediction at Multiple Timescales in V2P Conflicts Based on Gated Recurrent Unit

Improving pedestrian safety at urban intersections requires intelligent systems that should not only understand the actual vehicle–pedestrian (V2P) interaction state but also proactively anticipate the event’s future severity pattern. This paper presents a Gated Recurrent Unit-based system that aims to predict, up to 3 s ahead in time, the severity level of V2P encounters, depending on the current scene representation drawn from on-board radars’ data. A car-driving simulator experiment has been designed to collect sequential mobility features on a cohort of 65 licensed university students who faced different V2P conflicts on a planned urban route. To accurately describe the pedestrian safety condition during the encounter process, a combination of surrogate safety indicators, namely TAdv (Time Advantage) and T2 (Nearness of the Encroachment), are considered for modeling. Due to the nature of these indicators, multiple recurrent neural networks are trained to separately predict T2 continuous values and TAdv categories. Afterwards, their predictions are exploited to label serious conflict interactions. As a comparison, an additional Gated Recurrent Unit (GRU) neural network is developed to directly predict the severity level of inner-city encounters. The latter neural model reaches the best performance on the test set, scoring a recall value of 0.899. Based on selected threshold values, the presented models can be used to label pedestrians near accident events and to enhance existing intelligent driving systems.

A comparative study of multiple neural network for detection of COVID-19 on chest X-ray

Coronavirus disease of 2019 or COVID-19 is a rapidly spreading viral infection that has affected millions all over the world. With its rapid spread and increasing numbers, it is becoming overwhelming for the healthcare workers to rapidly diagnose the condition and contain it from spreading. Hence it has become a necessity to automate the diagnostic procedure. This will improve the work efficiency as well as keep the healthcare workers safe from getting exposed to the virus. Medical image analysis is one of the rising research areas that can tackle this issue with higher accuracy. This paper conducts a comparative study of the use of the recent deep learning models (VGG16, VGG19, DenseNet121, Inception-ResNet-V2, InceptionV3, Resnet50, and Xception) to deal with the detection and classification of coronavirus pneumonia from pneumonia cases. This study uses 7165 chest X-ray images of COVID-19 (1536) and pneumonia (5629) patients. Confusion metrics and performance metrics were used to analyze each model. Results show DenseNet121 (99.48% of accuracy) showed better performance when compared with the other models in this study.

Digital Hologram Watermarking Based on Multiple Deep Neural Networks Training Reconstruction and Attack

This paper proposes a method to embed and extract a watermark on a digital hologram using a deep neural network. The entire algorithm for watermarking digital holograms consists of three sub-networks. For the robustness of watermarking, an attack simulation is inserted inside the deep neural network. By including attack simulation and holographic reconstruction in the network, the deep neural network for watermarking can simultaneously train invisibility and robustness. We propose a network training method using hologram and reconstruction. After training the proposed network, we analyze the robustness of each attack and perform re-training according to this result to propose a method to improve the robustness. We quantitatively evaluate the results of robustness against various attacks and show the reliability of the proposed technique.

Modelling and intelligent prediction of sound quality for high-voltage fans on fuel cell vehicles

Since fuel cell vehicles have much higher heat dissipation requirements, 350-V high-voltage fans are adopted instead of traditional 12-V cooling fans, generating more aerodynamic noise. The installed fans are required to possess not only low sound pressure level but also good psychoacoustic performance. This paper is aimed at solving the complex correlation between subjective sound quality evaluation results and objective psychoacoustics parameters and establishing a sound quality prediction model for high-voltage fans. The noise signals of two high-voltage fans operating on a fuel cell vehicle under different running conditions are collected by an artificial head and preprocessed to acquire seven objective parameters. Then the subjective evaluation experiment on the annoyance of the noise samples is carried out based on pair-wise comparison method. A sample group of 23 adults is selected and a graphical user interface is programmed for test guiding. The subjective annoyance scores of the noise signals are obtained after data processing and effectiveness verification. By analyzing the tested results, the correlations between the subjective score and each of the single psychoacoustic parameters are summarized. Two sound quality prediction models are established by multiple linear regression and backpropagation neural network respectively, and the training results of the two methods are verified and compared, proving the reliability of neural network training results. With the established models, the sound quality of the high-voltage fans can be estimated effectively without complex subjective tests, contributing to improving the acoustic performance of fan products.

Learning Order Parameters from Videos of Skyrmion Dynamical Phases with Neural Networks

The ability to recognize dynamical processes in physical events and to abstract physical concepts or even to reveal physical laws, lies at the core of human cognitive development. Modern machine learning methods have been given high hopes for realizing such ability, but are still in their infancy. Here we present our progress in this capability. The main purposes of this paper are to use neural networks for classifying the dynamical skyrmion phases of videos and to demonstrate that neural networks can learn physical concepts from the dynamical process. To this end, we employ multiple neural networks to recognize the static phases (image format) and dynamical phases (video format) of a particle-based skyrmion model. Our results show that neural networks, without any prior knowledge, can not only correctly classify these phases, but also predict the phase boundaries that agree with those obtained by simulation. We further propose a parameter visualization scheme to interpret what neural networks have learned. We show that neural networks can learn two order parameters from videos of dynamical phases and predict the critical values of two order parameters. Finally, we demonstrate that only two order parameters are needed to identify videos of skyrmion dynamical phases. Hence, this parameter visualization scheme can be used to determine how many order parameters are needed to fully recognize the input phases. Our work sheds light on the future use of neural networks in discovering meaningful physical concepts and revealing unknown yet physical laws from videos.

POWERED LANDING GUIDANCE ALGORITHMS USING REINFORCEMENT LEARNING METHODS FOR LUNAR LANDER CASE

Any future planetary landing missions, just as demonstrated by Perseverance in 2021 Mars landing missions require advanced guidance, navigation, and control algorithms for the powered landing phase of the spacecraft to touch down a designated target with pinpoint accuracy (circular error precision < 5 m radius). This requires a landing system capable to estimate the craft’s states and map them to certain thrust commands for each craft’s engine. Reinforcement learning theory is used as an approach to manage the mapping guidance algorithm and translate it to engine thrust control commands. This work compares several reinforcement-learning-based approaches for a powered landing problem of a spacecraft in a two-dimensional (2-D) environment and identifies their advantages or disadvantages. Three methods in reinforcement learning, namely Q-Learning, and its extensions such as DQN and DDQN. They are benchmarked in terms of rewards and training time needed to land the Lunar Lander. It is found that the Q-Learning method also called Heuristic produced the highest efficiency. Another contribution of this paper is to show the combination usage of online weights � between the action selection process and action evaluation process, yields a higher reward, instead of separating them, which significantly enhances their optimization performance. The simulations of the powered guidance performance in a 2-D simulation environment highlight the effectiveness of DQN to handle multiple neural networks better than DDQN.

Hyperspectral anomaly detection based on the distinguishing features of a redundant difference-value network

ABSTRACT Hyperspectral anomaly detection is a key technique of unsupervised target detection. In the hyperspectral anomaly detection based on spectral dimensional transformation, the feature projection makes it easy to distinguish the ground objects which are not distinguishable in the original feature space. Although the means of spectral dimensional transformation can improve the distinguishable between diverse categories, it cannot highlight the anomalous targets. To be able to highlight anomalous targets while improving the diversity between different ground objects, an unsupervised network model of redundant difference-value network (RDVN) is proposed and applied to hyperspectral anomaly detection. RDVN is composed of multiple single-layer neural networks with the same structure and hyper-parameters. A group of training samples is used as the input of the networks, and the difference between the activation values of any network and benchmark network is used as the error for backpropagation. After the training is completed, the difference-value between the activation values of the two networks is used as a distinguishing feature (DF). Finally, DF is used as the input of the anomaly detector to obtain the detection results. Experimental results demonstrate that the proposed algorithm can achieve higher detection accuracy. DF not only highlights the anomalous target to increase the true positive rate but also increases the discriminability between different categories, thereby reducing the false-positive rate.

Learning to Recognize Irregular Features on Leather Surfaces

As part of industrial quality control in the leather industry, it is important to identify the abnormal features in wet-blue leather samples. Manual inspection of leather samples is the current norm in industrial settings. To comply with the current industrial standards that advocate large-scale automation, visual inspection based leather processing is imperative. Visual inspection of irregular surfaces is a challenging problem as the characteristics of the abnormalities can take a variety of shape and color variations. The aim of this work is to automatically categorize leather images into normal or abnormal by visual analysis of the surfaces. To achieve this aim, a deep learning based approach is devised that learns to recognize regular and irregular leather surfaces and categorize leather images on its basis. To this end, we propose an ensemble of multiple convolutional neural networks for classifying leather images. The proposed ensemble network exhibited competitive performance obtaining 92.68% test accuracy on our own curated leather images dataset.

Longitudinal, transverse and ultrasound vibration for the prediction of stiffness using models incorporating features in Pinus sylvestris timber

Non-destructive testing was used to predict the static modulus of elasticity (MOES) of Scots pine (Pinus sylvestris) timber from the northeast of Spain. Three vibration tests were performed, longitudinal, flatwise and edgewise, to obtain the dynamic modulus of elasticity (MOEdyn) based on the fundamental resonant frequencies. The MOEdyn was additionally obtained from ultrasound tests. Measurements of different features were performed of the various samples, which were also subjected to a bending test to find the MOES. Different types of models, simple linear regression (SLR), multiple linear regression (MLR) and artificial neural network (ANN), were generated to predict the MOES based on the study variables. The predictive capacity of the different models was analysed by comparing the root mean square error (RMSE) obtained using the tenfold cross-validation method. The vibration techniques showed a better MOES prediction than the ultrasound techniques. The MOEdyn obtained from the fundamental resonant frequency of the edgewise flexural vibration (MOEEV) was the variable that best predicted the MOES. The error of the SLR with MOEEV was not significantly improved by any other model, whether univariate or multivariate. The ANN-based models did not significantly improve the error of the MLR-based models.

ECG beat classification using neural classifier based on deep autoencoder and decomposition techniques

This paper proposes an ECG beat classification system based on deep autoencoder as feature extractor and a system of multiple neural networks as classifier. The objectives are as follows: First is simplifying the feature extraction step by applying the deep autoencoder, which permits defining high level features without neither pre-processing stage nor expert intervention. Second is enhancing the classification performance by decomposing the original multi-class problem into simpler binary subproblems and solving them using independent classifiers. Third is overcoming the problem of imbalanced data, by applying an oversampling method after the decomposition of the original problem. This allows adding synthetic samples according the number of training instances in each subproblem. To evaluate the proposed system, we conduct experiments on MIT-BIH arrhythmia dataset and we consider the recommendations of the Association for the Advancement of Medical Instrumentation, which defines five classes of interest. Furthermore, we perform two types of tests, i.e. intra- and inter-patient, and compare the obtained results with some of the state-of-the-art methods. We show that solving each subproblem independently can enhance the accuracy, sensitivity and specificity.

An Analytical Study On Forecasting Exchange Rate In The Philippines Using Multi-Layer Feed Forward Neural Network

Exchange Rate is one of the economic indicators in the Philippines. It is the value of the nation’s currency versus the currency of another country or economic zone. This study aims to forecast the monthly Exchange Rate (y) of the Philippines from November 2018 to December 2023 using Multiple Linear Regression and Multi-Layer Feed Forward Neural Network. The researchers investigate the behaviour of each independent variables – Inflation Rate (x1), Balance of Payments (x2), Interest Rate (x3), Producer’s Price Index (x4), Export (x5), Import (x6), Money Supply (x7), and Consumer’s Price Index (x8) from Philippine Statistics Authority (PSA) starts from January 2007 up to October 2018. Multiple Linear Regression (MLR) was used to identify significant predictors among these independent variables. The Exchange Rate (y) had undergone first difference transformation. Upon running the regression analysis, it has concluded that only two independent variables are significant predictors, namely: Balance of Payments (x2) and Import (x6). Through these significant predictors, the MLR model was formulated. On the other hand, Multi-Layer Feed forward Neural Network (MFFNN) was also used to determine the forecasted values of Exchange Rate (y) for the next five years (2018-2023) given the said independent variables and obtained a model. The researchers compared the model of Multiple Linear Regression and Multi-Layer Feed Forward Neural Network by evaluating the forecasting accuracy of each method.It was concluded that Multi-Layer Feed forward Neural Network is the best fitting model for forecasting the&#x0D; Exchange rate (y) in the Philippines. This paper will serve as a tool of awareness for the government to forsee the trend of Exchange Rate in the Philippines on the next five years (2018-2023) for Monetary Policy making and to prevent the possible depreciation of peso vs. dollar.

An Analysis on Forecasting Inflation Rate in the Philippines: A Recurrent Neural Network Method Approach

Inflation rate is the proceeding rise within the common level of costs of products and services in an economy over a certain span of time. In 2018, the Philippines has the highest inflation rate among the 10 South East Asian countries. The objective of this research is to forecast the inflation rate of the Philippines for the next five years (2019-2023). Also, the researchers compared the results obtained from the Multiple Linear Regression and Recurrent Neural Network (RNN) performed in MATLAB to determine which of these two models will be the better model in forecasting inflation rate. In this study, the researchers observed the behavior of the Inflation Rate(y) and its economic factors such as Import (x1), Export (x2), Money Supply (x3), Gross Domestic Product (x4), Gross National Product (x5), Expenditure (x6) and Exchange Rate (x7). Using Multiple Linear Regression, this study determined that the significant predictors are Money Supply (x3) and Expenditures (x6). By evaluating the forecast efficiency of the two methods, the researchers concluded that Multilayered Recurrent Neural Network outperforms Multiple Linear Regression in predicting inflation rate of the Philippines. This paper can be useful to the Philippine Government on their decisions about monetary policy making since forecasting the inflation rate has a huge importance and impact in conducting monetary policy.

Psychiatric Neural Networks and Precision Therapeutics by Machine Learning.

Learning and environmental adaptation increase the likelihood of survival and improve the quality of life. However, it is often difficult to judge optimal behaviors in real life due to highly complex social dynamics and environment. Consequentially, many different brain regions and neuronal circuits are involved in decision-making. Many neurobiological studies on decision-making show that behaviors are chosen through coordination among multiple neural network systems, each implementing a distinct set of computational algorithms. Although these processes are commonly abnormal in neurological and psychiatric disorders, the underlying causes remain incompletely elucidated. Machine learning approaches with multidimensional data sets have the potential to not only pathologically redefine mental illnesses but also better improve therapeutic outcomes than DSM/ICD diagnoses. Furthermore, measurable endophenotypes could allow for early disease detection, prognosis, and optimal treatment regime for individuals. In this review, decision-making in real life and psychiatric disorders and the applications of machine learning in brain imaging studies on psychiatric disorders are summarized, and considerations for the future clinical translation are outlined. This review also aims to introduce clinicians, scientists, and engineers to the opportunities and challenges in bringing artificial intelligence into psychiatric practice.

Inhibitory control in children with tic disorder: aberrant fronto-parietal network activity and connectivity

Chronic tic disorders, including Tourette syndrome, are typically thought to have deficits in cognitive inhibition and top down cognitive control due to the frequent and repetitive occurrence of tics, yet studies reporting task performance results have been equivocal. Despite similar behavioural performance, individuals with chronic tic disorder have exhibited aberrant patterns of neural activation in multiple frontal and parietal regions relative to healthy controls during inhibitory control paradigms. In addition to these top down attentional control regions, widespread alterations in brain activity across multiple neural networks have been reported. There is a dearth, however, of studies examining event-related connectivity during cognitive inhibitory paradigms among affected individuals. The goal of this study was to characterize neural oscillatory activity and effective connectivity, using a case–control design, among children with and without chronic tic disorder during performance of a cognitive inhibition task. Electroencephalogram data were recorded in a cohort of children aged 8–12 years old (60 with chronic tic disorder, 35 typically developing controls) while they performed a flanker task. While task accuracy did not differ by diagnosis, children with chronic tic disorder displayed significant cortical source-level, event-related spectral power differences during incongruent flanker trials, which required inhibitory control. Specifically, attenuated broad band oscillatory power modulation within the anterior cingulate cortex was observed relative to controls. Whole brain effective connectivity analyses indicated that children with chronic tic disorder exhibit greater information flow between the anterior cingulate and other fronto-parietal network hubs (midcingulate cortex and precuneus) relative to controls, who instead showed stronger connectivity between central and posterior nodes. Spectral power within the anterior cingulate was not significantly correlated with any connectivity edges, suggesting lower power and higher connectivity are independent (versus resultant) neural mechanisms. Significant correlations between clinical features, task performance and anterior cingulate spectral power and connectivity suggest this region is associated with tic impairment (r = −0.31, P = 0.03) and flanker task incongruent trial accuracy (r’s = −0.27 to −0.42, P’s = 0.0008–0.04). Attenuated activation of the anterior cingulate along with dysregulated information flow between and among nodes within the fronto-parietal attention network may be neural adaptations that result from frequent engagement of neural pathways needed for inhibitory control in chronic tic disorder.

Combine and conquer: event reconstruction with Bayesian Ensemble Neural Networks

Ensemble learning is a technique where multiple component learners are combined through a protocol. We propose an Ensemble Neural Network (ENN) that uses the combined latent-feature space of multiple neural network classifiers to improve the representation of the network hypothesis. We apply this approach to construct an ENN from Convolutional and Recurrent Neural Networks to discriminate top-quark jets from QCD jets. Such ENN provides the flexibility to improve the classification beyond simple prediction combining methods by linking different sources of error correlations, hence improving the representation between data and hypothesis. In combination with Bayesian techniques, we show that it can reduce epistemic uncertainties and the entropy of the hypothesis by simultaneously exploiting various kinematic correlations of the system, which also makes the network less susceptible to a limitation in training sample size.

Superposition learning-based model for prediction of E.coli in groundwater using physico-chemical water quality parameters

The prediction of waterborne bacteria is crucial to prevent health risks. Therefore, there is a need to study the quality of groundwater by predicting the presence of E.coli. The experimental data for prediction is obtained from BITS-UVA (University of Virginia) groundwater contamination project, having 1301 experimental laboratory results that will be synthesized to test the physical, chemical, and microbiological parameters of water. In this study, a superposition-based learning algorithm (SLA) is proposed to observe the patterns of ANN-based sensitivity analysis to determine the importance of each water quality parameter resulting in the prediction of E.coli in groundwater. Mean Square Error (MSE) and the Coefficient of determination (R2) are calculated using MATLAB (R2019b, Mathworks, Natik, MA, USA) software for model performance evaluation. The highest correlation is observed between E.coli and the pH values, whereas the lowest correlation is observed with Dissolved Oxygen. In order to find out the uncertainty in the output of our mathematical model, a sensitivity analysis for the seven models is carried out. The results show that the model having Turbidity, pH, Total Dissolved Salts (TDS), and Electrical Conductivity as inputs displayed the best performance. The model architecture constitutes three hidden layers, twenty neurons in each layer, which is optimized using the Bayesian Regularization training algorithm (BR) with the overall highest R2 values of 0.90 and lowest MSE values of 0.0892. Patterns of the trained neural network are presented in superposition. After training, it can be concluded that the superposition models based on Grover's algorithm is more efficient in predicting all patterns in the counts of E.coli in groundwater. The algorithm is superimposed on multiple neural network architectures and returns a trained neural network. In addition to accurate results, there is also a need to automate the process of real-time bacterial monitoring for minimizing the error.

Towards data-free gating of heterogeneous pre-trained neural networks

The combination and aggregation of knowledge from multiple neural networks can be commonly seen in the form of mixtures of experts. However, such combinations are usually done using networks trained on the same tasks, with little mention of the combination of heterogeneous pre-trained networks, especially in the data-free regime. This paper proposes multiple data-free methods for the combination of heterogeneous neural networks, ranging from the utilization of simple output logit statistics, to training specialized gating networks. The gating networks decide whether specific inputs belong to specific networks based on the nature of the expert activations generated. The experiments revealed that the gating networks, including the universal gating approach, constituted the most accurate approach, and therefore represent a pragmatic step towards applications with heterogeneous mixtures of experts in a data-free regime. The code for this project is hosted on github at https://github.com/cwkang1998/network-merging.

Moisture-carryover performance optimization using physics-constrained machine learning

A data-driven model for predicting moisture carryover (MCO) in the General Electric Type-4 boiling water reactor (BWR) was constructed using a physics-constrained artificial intelligence technique. An accurate prediction of the MCO is of great value for commercial BWR operators as it can be used to modify the operational plan during a power cycle to mitigate high MCO, thereby avoiding elevated dose to on-site personnel and damage to turbine components. Using data from operational plants and preliminary features selected through physics and engineering analyses, a neural network based model for predicting MCO was built. A final feature set was then obtained through a hyperspace optimization performed using a genetic algorithm. Multiple neural network models possessing good generalization were obtained, the best of these having a mean-square error (MSE) of 9.69E-5 for prediction of an un-seen cycle, which is in agreement with the uncertainty in the measured MCO data. This predictive capability is of great value for the planning of a power generation cycle, and for scheduling of operations for cycles already underway.