Recurrent Neural Network Learning Research Articles

Prediction of the position of external markers using a recurrent neural network trained with unbiased online recurrent optimization for safe lung cancer radiotherapy

Background and Objective: During lung cancer radiotherapy, the position of infrared reflective objects on the chest can be recorded to estimate the tumor location. However, radiotherapy systems have a latency inherent to robot control limitations that impedes the radiation delivery precision. Prediction with online learning of recurrent neural networks (RNN) allows for adaptation to non-stationary respiratory signals, but classical methods such as real-time recurrent learning (RTRL) and truncated backpropagation through time are respectively slow and biased. This study investigates the capabilities of unbiased online recurrent optimization (UORO) to forecast respiratory motion and enhance safety in lung radiotherapy. Methods: We used nine observation records of the three-dimensional (3D) position of three external markers on the chest and abdomen of healthy individuals breathing during intervals from 73s to 222s. The sampling frequency was 10Hz, and the amplitudes of the recorded trajectories range from 6mm to 40mm in the superior-inferior direction. We forecast the 3D location of each marker simultaneously with a horizon value (the time interval in advance for which the prediction is made) between 0.1s and 2.0s, using an RNN trained with UORO. We compare its performance with an RNN trained with RTRL, least mean squares (LMS), and offline linear regression. We provide closed-form expressions for quantities involved in the gradient loss calculation in UORO, thereby making its implementation efficient. Training and cross-validation were performed during the first minute of each sequence. Results: On average over the horizon values considered and the nine sequences, UORO achieves the lowest root-mean-square (RMS) error and maximum error among the compared algorithms. These errors are respectively equal to 1.3mm and 8.8mm, and the prediction time per time step was lower than 2.8ms (Dell Intel core i9-9900K 3.60 GHz). Linear regression has the lowest RMS error for the horizon values 0.1s and 0.2s, followed by LMS for horizon values between 0.3s and 0.5s, and UORO for horizon values greater than 0.6s. Conclusions: UORO can accurately predict the 3D position of external markers for intermediate to high response times with an acceptable time performance. This will help limit unwanted damage to healthy tissues caused by radiotherapy.

Continual Sequence Modeling With Predictive Coding.

Recurrent neural networks (RNNs) have been proved very successful at modeling sequential data such as language or motions. However, these successes rely on the use of the backpropagation through time (BPTT) algorithm, batch training, and the hypothesis that all the training data are available at the same time. In contrast, the field of developmental robotics aims at uncovering lifelong learning mechanisms that could allow embodied machines to learn and stabilize knowledge in continuously evolving environments. In this article, we investigate different RNN designs and learning methods, that we evaluate in a continual learning setting. The generative modeling task consists in learning to generate 20 continuous trajectories that are presented sequentially to the learning algorithms. Each method is evaluated according to the average prediction error over the 20 trajectories obtained after complete training. This study focuses on learning algorithms with low memory requirements, that do not need to store past information to update their parameters. Our experiments identify two approaches especially fit for this task: conceptors and predictive coding. We suggest combining these two mechanisms into a new proposed model that we label PC-Conceptors that outperforms the other methods presented in this study.

Current State and Future Directions for Learning in Biological Recurrent Neural Networks: A Perspective Piece

This perspective piece came about through the Generative Adversarial Collaboration (GAC) series of workshops organized by the Computational Cognitive Neuroscience (CCN) conference in 2020. We brought together a number of experts from the field of theoretical neuroscience to debate emerging issues in our understanding of how learning is implemented in biological recurrent neural networks. Here, we will give a brief review of the common assumptions about biological learning and the corresponding findings from experimental neuroscience and contrast them with the efficiency of gradient-based learning in recurrent neural networks commonly used in artificial intelligence. We will then outline the key issues discussed in the workshop: synaptic plasticity, neural circuits, theory-experiment divide, and objective functions. Finally, we conclude with recommendations for both theoretical and experimental neuroscientists when designing new studies that could help to bring clarity to these issues.

Deep Learning Neural Network Prediction System Enhanced with Best Window Size in Sliding Window Algorithm for Predicting Domestic Power Consumption in a Residential Building.

Buildings are considered to be one of the world's largest consumers of energy. The productive utilization of energy will spare the accessible energy assets for the following ages. In this paper, we analyze and predict the domestic electric power consumption of a single residential building, implementing deep learning approach (LSTM and CNN). In these models, a novel feature is proposed, the “best N window size” that will focus on identifying the reliable time period in the past data, which yields an optimal prediction model for domestic energy consumption known as deep learning recurrent neural network prediction system with improved sliding window algorithm. The proposed prediction system is tuned to achieve high accuracy based on various hyperparameters. This work performs a comparative study of different variations of the deep learning model and records the best Root Mean Square Error value compared to other learning models for the benchmark energy consumption dataset.

Multi-step-ahead prediction of river flow using NARX neural networks and deep learning LSTM

Abstract Kelantan river (Sungai Kelantan in Malaysia) basin is one of the essential catchments as it has a history of flood events. Numerous studies have been conducted in river basin modelling for the prediction of flow and mitigation of flooding events as well as water resource management. Therefore, having multi-step-ahead forecasting for river flow (RF) is of important research interest in this regard. This study presents four different approaches for multi-step-ahead forecasting for the Kelantan RF, using NARX (nonlinear autoregressive with exogenous inputs) neural networks and deep learning recurrent neural networks called LSTM (long short-term memory). The dataset used was obtained in monthly record for 29 years between January 1988 and December 2016. The results show that two recursive methods using NARX and LSTM are able to do multi-step-ahead forecasting on 52 series of test datasets with NSE (Nash–Sutcliffe efficiency coefficient) values of 0.44 and 0.59 for NARX and LSTM, respectively. For few-step-ahead forecasting, LSTM with direct sequence-to-sequence produces promising results with a good NSE value of 0.75 (in case of two-step-ahead forecasting). However, it needs a larger data size to have better performance in longer-step-ahead forecasting. Compared with other studies, the data used in this study is much smaller.

Telemetry Data Compression Algorithm Using Balanced Recurrent Neural Network and Deep Learning.

Telemetric information is great in size, requiring extra room and transmission time. There is a significant obstruction of storing or sending telemetric information. Lossless data compression (LDC) algorithms have evolved to process telemetric data effectively and efficiently with a high compression ratio and a short processing time. Telemetric information can be packed to control the extra room and association data transmission. In spite of the fact that different examinations on the pressure of telemetric information have been conducted, the idea of telemetric information makes pressure incredibly troublesome. The purpose of this study is to offer a subsampled and balanced recurrent neural lossless data compression (SB-RNLDC) approach for increasing the compression rate while decreasing the compression time. This is accomplished through the development of two models: one for subsampled averaged telemetry data preprocessing and another for BRN-LDC. Subsampling and averaging are conducted at the preprocessing stage using an adjustable sampling factor. A balanced compression interval (BCI) is used to encode the data depending on the probability measurement during the LDC stage. The aim of this research work is to compare differential compression techniques directly. The final output demonstrates that the balancing-based LDC can reduce compression time and finally improve dependability. The final experimental results show that the model proposed can enhance the computing capabilities in data compression compared to the existing methodologies.

Recurrent neural networks in cyber security

The recent increase in cyber-attacks and malware clearly demonstrates that current countermeasures do not seem to be enough to protect against it, as hackers become more cautious in their approach with the cunning of developing systems that automatically rewrite and reorder their malicious software to avoid detection. Typical machine learning approaches that learn a classifier based on a manually created feature vector are not robust enough to such reordering. Hence, the need for an effective automated cyber security solution using deep neural networks. In this article, we demonstrate research on the effectiveness of recurrent neural networks (RNNs) for combat in cyberspace. The conducted experiment shows that RNN with Long Short Term Memory (LSTM) performs much better than classical machine learning algorithms (SVM and Random Forest) with accuracy of 99.70%, 98.55% and 99.42%, respectively. This is possible because RNNs have a built-in memory that can remember multiple previous states and implicitly extract distinctive features, hidden complex structure, and complex sequential relationships in the data, which helps achieve better accuracy. This paper describes an investigation of a recurrent neural network (RNN) model for cyber security using domain malware detection as an application area. The performance of RNN and other classical machine learning classifiers is studied and evaluated for malware classification in the cyber security usage domain and compared. From this study, it can be seen that RNN has better accuracy than classical machine learning classifiers (SVM and Random forest). This is possible because RNNs have a built-in memory capability that can store and replay multiple previous states, and implicitly extract salient features, hidden/underlying complex structure, and complex sequential relationships in the data, which help achieve better accuracy. Thus, it will be useful for creating a real-time application for analyzing malicious activities on the network.

A Novel Learning Model of Driver Fatigue Features Representation for Steering Wheel Angle

In the field of advanced driver assistance systems (ADAS), effective learning of driver fatigue characteristics representation is a major challenge due to uncertainties of both real roads and drivers. To tackle this problem, this paper proposes a novel model of learning interpretable representations for fatigue features, so as to improve the performance of the monitoring system on real roads through the learned features. First, the approximate entropy of the Steering Wheel Angle (SWA)sequence is used to crop adaptive lengths of the Recurrent Neural Network (RNN) input data. Then, it will learn statistical indexes to discover the random steering characteristics. The Long Short-Term Memory (LSTM) unit can memorize drivers’ long-term operating characteristics and instantaneous change patterns, and mine their fatigue characteristics. Finally, the information gain method is used to discover the strong correlation between potential characteristics and fatigue levels, thereby obtaining the best feature representation for driver fatigue. The proposed method makes use of the advantages of recurrent neural network learning to explicitly capture various potential characteristics of fatigue driving and interactions between non-linear characteristics in driving. Experimental results on real driving data validate the effectiveness of our proposed method.

Abd-Rnn: Attention Based Bi-Directional Deep Learning Recurrent Neural Networks for Aquaculture Water Quality Prediction

Abd-Rnn: Attention Based Bi-Directional Deep Learning Recurrent Neural Networks for Aquaculture Water Quality Prediction

Evaluation of deep learning techniques for identification of sarcoma-causing carcinogenic mutations.

The abnormal growth of human healthy cells is called cancer. One of the major types of cancer is sarcoma, mostly found in human bones and soft tissue cells. It commonly occurs in children. According to a survey of the United States of America, there are more than 17,000 sarcoma patients registered each year which is 15% of all cancer cases. Recognition of cancer at its early stage saves many lives. The proposed study developed a framework for the early detection of human sarcoma cancer using deep learning Recurrent Neural Network (RNN) algorithms. The DNA of a human cell is made up of 25,000 to 30,000 genes. Each gene is represented by sequences of nucleotides. The nucleotides in a sequence of a driver gene can change which is termed as mutations. Some mutations can cause cancer. There are seven types of a gene whose mutation causes sarcoma cancer. The study uses the dataset which has been taken from more than 134 samples and includes 141 mutations in 8 driver genes. On these gene sequences RNN algorithms Long and Short-Term Memory (LSTM), Gated Recurrent Units and Bi-directional LSTM (Bi-LSTM) are used for training. Rigorous testing techniques such as Self-consistency testing, independent set testing, 10-fold cross-validation test are applied for the validation of results. These validation techniques yield several metrics such as Area Under the Curve (AUC), sensitivity, specificity, Mathew's correlation coefficient, loss, and accuracy. The proposed algorithm exhibits an accuracy of 99.6% with an AUC value of 1.00.

CRNNTL: Convolutional Recurrent Neural Network and Transfer Learning for QSAR Modeling in Organic Drug and Material Discovery.

Molecular latent representations, derived from autoencoders (AEs), have been widely used for drug or material discovery over the past couple of years. In particular, a variety of machine learning methods based on latent representations have shown excellent performance on quantitative structure–activity relationship (QSAR) modeling. However, the sequence feature of them has not been considered in most cases. In addition, data scarcity is still the main obstacle for deep learning strategies, especially for bioactivity datasets. In this study, we propose the convolutional recurrent neural network and transfer learning (CRNNTL) method inspired by the applications of polyphonic sound detection and electrocardiogram classification. Our model takes advantage of both convolutional and recurrent neural networks for feature extraction, as well as the data augmentation method. According to QSAR modeling on 27 datasets, CRNNTL can outperform or compete with state-of-art methods in both drug and material properties. In addition, the performances on one isomers-based dataset indicate that its excellent performance results from the improved ability in global feature extraction when the ability of the local one is maintained. Then, the transfer learning results show that CRNNTL can overcome data scarcity when choosing relative source datasets. Finally, the high versatility of our model is shown by using different latent representations as inputs from other types of AEs.

Ultrasound Elastography under Deep Learning Algorithm to Analyze the Therapeutic Effect of Clustered Regularly Interspaced Short Palindromic Repeats Short Hairpin Ribonucleic Acid Nanoparticles on Cervical Cancer.

This study aimed to analyze the effect of the deep learning algorithm on ultrasound elastography on the treatment of cervical cancer with clustered regularly interspaced short palindromic repeats (CRISPR) short hairpin ribonucleic acid (shRNA) nanoparticles, aiming to provide a reference for the clinical application of deep learning to analyze the therapeutic effect of the disease. In this study, CRISPR and shRNA plasmid nanoparticle drugs were used to treat 55 patients with cervical cancer in the experimental group, and normal saline was injected to another 53 patients in the control group, so compare the effect of nanoparticles in the treatment of cervical cancer. Professional doctors and the recurrent neural network (RNN) intelligent algorithm were used to score cervical cancer based on the ultrasound elastograph images by taking blue, green, and red (BGR) as diagnosis criteria. As a result, the experimental group had a total of 217 points before drug administration and a total of 224 points after drug administration. Each patient had an average increase of 0.13 points. The control group had a total of 200 points before drug administration and a total of 223 points after drug administration, and each patient had an average increase of 0.43 points. The experimental group was obviously different from the control group (P < 0.05). Each tissue image output by the RNN was clearer than the original image, and the score given by intelligent calculation was faster than that of professional doctors. The monitoring effect of the deep learning RNN intelligent algorithm on the therapeutic effect of nanomedicine was analyzed. It was found that the average accuracy of the experimental group and the control group was 98.95% and 90.34%, respectively; and the experimental group was greatly different from the control group (P < 0.05). In short, nano-CRISPR and shRNA drugs had remarkable effects on the treatment of cervical cancer, and the scores given by the deep learning intelligent algorithm were faster and more accurate, which provided theoretical guidance for the clinical application of deep learning algorithms to analyze the treatment effects of diseases.

Deep Learning Recurrent Neural Network for Concussion Classification in Adolescents Using Raw Electroencephalography Signals: Toward a Minimal Number of Sensors.

Artificial neural networks (ANNs) are showing increasing promise as decision support tools in medicine and particularly in neuroscience and neuroimaging. Recently, there has been increasing work on using neural networks to classify individuals with concussion using electroencephalography (EEG) data. However, to date the need for research grade equipment has limited the applications to clinical environments. We recently developed a deep learning long short-term memory (LSTM) based recurrent neural network to classify concussion using raw, resting state data using 64 EEG channels and achieved high accuracy in classifying concussion. Here, we report on our efforts to develop a clinically practical system using a minimal subset of EEG sensors. EEG data from 23 athletes who had suffered a sport-related concussion and 35 non-concussed, control athletes were used for this study. We tested and ranked each of the original 64 channels based on its contribution toward the concussion classification performed by the original LSTM network. The top scoring channels were used to train and test a network with the same architecture as the previously trained network. We found that with only six of the top scoring channels the classifier identified concussions with an accuracy of 94%. These results show that it is possible to classify concussion using raw, resting state data from a small number of EEG sensors, constituting a first step toward developing portable, easy to use EEG systems that can be used in a clinical setting.

Emotion identification by dynamic entropy and ensemble learning from electroencephalogram signals

AbstractEmotions are biologically based psychological states brought on by neurophysiologic changes, variously associated with thoughts, feelings, behavioral responses, and a degree of pleasure or displeasure. In order to assure active collaboration and trigger appropriate emotional input, accurate identification of human emotions is important. Poor generalization capacity induced by individual variations in emotion perceptions is indeed a concern in the current methods of emotion recognition. This work proposes a new dynamic pattern learning system based on entropy to allow electroencephalogram (EEG) signals for subject‐independent emotion recognition with strong generalization and classification through Recurrent Neural Network and Ensemble learning. First, dynamic entropy measurements are used to derive consecutive entropy values over time from EEG signals in quantitative EEG calculations. Experiment findings indicate that in order to distinguish negative and positive emotions, the highest average accuracy of 94.67% is achieved. In addition, the findings have completely shown that this approach produces outstanding performance for emotion detection across individuals relative to recent studies.

Online Machine Learning of Available Capacity for Vehicle-to-Grid Services during the Coronavirus Pandemic

Vehicle-to-grid services make use of the aggregated capacity available from a fleet of vehicles to participate in energy markets, help integrate renewable energy in the grid and balance energy use. In this paper, the critical components of such a service are described in the context of a commercial service that is currently under development. Key among these components is the prediction of available capacity at a future time. In this paper, we extend a previous work that used a deep learning recurrent neural network for this task to include online machine learning, which enables the network to continually refine its predictions based on observed behaviour. The coronavirus pandemic that was declared in 2020 resulted in closures of the university and substantial changes to the behaviour of the university fleet. In this work, the impact of this change in vehicles usage was used to test the predictions of a network initially trained using vehicle trip data from 2019 with and without online machine learning. It is shown that prediction error is significantly reduced using online machine learning, and it is concluded that a similar capability will be of critical importance for a commercial service such as the one described in this paper.

Forecasting emergency department admissions

The emergency department of a hospital plays an extremely important role in the healthcare of patients. To maintain a high quality service, clinical professionals need information on how patient flow will evolve in the immediate future. With accurate emergency department forecasts it is possible to better manage available human resources by allocating clinical staff before peak periods, thus preventing service congestion, or releasing clinical staff at less busy times. This paper describes a solution developed for the presentation of hourly, four-hour, eight-hour and daily number of admissions to a hospital’s emergency department. A 10-year history (2009–2018) of the number of emergency admissions in a Portuguese hospital was used. To create the models several methods were tested, including exponential smoothing, SARIMA, autoregressive and recurrent neural network, XGBoost and ensemble learning. The models that generated the most accurate hourly time predictions were the recurrent neural network with one-layer (sMAPE = 23.26%) and with three layers (sMAPE = 23.12%) and XGBoost (sMAPE = 23.70%). In terms of efficiency, the XGBoost method has by far outperformed all others. The success of the recurrent neuronal network and XGBoost machine learning methods applied to the prediction of the number of emergency department admissions has been demonstrated here, with an accuracy that surpasses the models found in the literature.

Recurrent Neural Network and Reinforcement Learning Model for COVID-19 Prediction.

Detection and prediction of the novel Coronavirus present new challenges for the medical research community due to its widespread across the globe. Methods driven by Artificial Intelligence can help predict specific parameters, hazards, and outcomes of such a pandemic. Recently, deep learning-based approaches have proven a novel opportunity to determine various difficulties in prediction. In this work, two learning algorithms, namely deep learning and reinforcement learning, were developed to forecast COVID-19. This article constructs a model using Recurrent Neural Networks (RNN), particularly the Modified Long Short-Term Memory (MLSTM) model, to forecast the count of newly affected individuals, losses, and cures in the following few days. This study also suggests deep learning reinforcement to optimize COVID-19's predictive outcome based on symptoms. Real-world data was utilized to analyze the success of the suggested system. The findings show that the established approach promises prognosticating outcomes concerning the current COVID-19 pandemic and outperformed the Long Short-Term Memory (LSTM) model and the Machine Learning model, Logistic Regresion (LR) in terms of error rate.

Variable weight algorithm for convolutional neural networks and its applications to classification of seizure phases and types

Deep learning techniques have recently achieved impressive results and raised expectations in the domains of medical diagnosis and physiological signal processing. The widely adopted methods include convolutional neural networks (CNNs) and recurrent neural networks (RNNs). However, the existing models possess static connection weights between layers, which might limit the generalization capability and the classification performance of the models as the weights of different layers are fixed after training. Furthermore, to deal with a large amount of data, a neural network with a sufficiently large size is required. This paper proposes the variable weight convolutional neural networks (VWCNNs), which are a type of network structure employing dynamic weights instead of static weights in their convolutional layers and fully-connected layers. VWCNNs are able to adapt to different characteristics of input data and can be viewed as an infinite number of traditional, fixed-weight CNNs. We will show that the proposed VWCNN structure outperforms the conventional CNN in terms of the classification accuracy, generalization capability, and robustness when the inputs are contaminated by noise. In this paper, VWCNNs are applied to the classification of three seizure phases (seizure-free, pre-seizure and seizure) based on measured electroencephalography (EEG) data. VWCNNs achieve 100% test accuracy and show strong robustness in the classification of the three seizure phases, and thus show the potential to be a useful classification tool for medical diagnosis. Furthermore, the classification of seven types of seizures is investigated in this paper using the world’s largest open source database of seizure recordings, TUH EEG seizure corpus. Comparisons with conventional CNNs, RNN, MobileNet, ResNet, DenseNet and traditional machine learning methods including random forest, decision tree, support vector machine, K-nearest neighbours, standard neural networks, and Naïve Bayes are being conducted using realistic test data sets. The results demonstrate that VWCNNs have advantages over other classifiers in terms of classification accuracy and robustness.

Dynamic selective auditory attention detection using RNN and reinforcement learning

The cocktail party phenomenon describes the ability of the human brain to focus auditory attention on a particular stimulus while ignoring other acoustic events. Selective auditory attention detection (SAAD) is an important issue in the development of brain-computer interface systems and cocktail party processors. This paper proposes a new dynamic attention detection system to process the temporal evolution of the input signal. The proposed dynamic SAAD is modeled as a sequential decision-making problem, which is solved by recurrent neural network (RNN) and reinforcement learning methods of Q-learning and deep Q-learning. Among different dynamic learning approaches, the evaluation results show that the deep Q-learning approach with RNN as agent provides the highest classification accuracy (94.2%) with the least detection delay. The proposed SAAD system is advantageous, in the sense that the detection of attention is performed dynamically for the sequential inputs. Also, the system has the potential to be used in scenarios, where the attention of the listener might be switched in time in the presence of various acoustic events.

Fractional order PID-based adaptive fault-tolerant cooperative control of networked unmanned aerial vehicles against actuator faults and wind effects with hardware-in-the-loop experimental validation

This paper proposes an adaptive fault-tolerant cooperative control (FTCC) scheme for networked unmanned aerial vehicles (UAVs) in the presence of actuator faults and wind effects by artfully introducing fractional order calculus, proportional–integral–derivative (PID), and recurrent neural networks. Fractional order sliding-mode surface and PID-type error mapping are first utilized to transform the synchronization tracking errors of all UAVs into a new set of errors. Then, based on these newly constructed errors, an FTCC scheme is developed to synchronously track their references. Moreover, Butterworth low-pass filter (BLF) and recurrent neural network (RNN) learning strategies are assimilated to handle the unknown terms induced by the actuator faults and wind effects. Finally, theoretical analysis and comparative hardware-in-the-loop experimental demonstrations have shown the effectiveness of the proposed control scheme.