LSTM-based Recurrent Neural Network Research Articles

Heteroscedasticity effects as component to future stock market predictions using RNN-based models.

Heteroscedasticity effects are useful for forecasting future stock return volatility. Stock volatility forecasting provides business insight into the stock market, making it valuable information for investors and traders. Predicting stock volatility is a crucial task and challenging. This study proposes a hybrid model that predicts future stock volatility values by considering the heteroscedasticity element of the stock price. The proposed model is a combination of Generalized Autoregressive Conditional Heteroskedasticity (GARCH) and a well-known Recurrent Neural Network (RNN) algorithm Long Short-Term Memory (LSTM). This proposed model is referred to as GARCH-LSTM model. The proposed model is expected to improve prediction accuracy by considering heteroscedasticity elements. First, the GARCH model is employed to estimate the model parameters. After that, the ARCH effect test is used to test the residuals obtained from the model. Any untrained heteroscedasticity element must be found using this step. The hypothesis of the ARCH test yielded a p-value less than 0.05 indicating there is valuable information remaining in the residual, known as heteroscedasticity element. Next, the dataset with heteroscedasticity is then modelled using an LSTM-based RNN algorithm. Experimental results revealed that hybrid GARCH-LSTM had the lowest MAE (7.961), RMSE (10.466), MAPE (0.516) and HMAE (0.005) values compared with a single LSTM. The accuracy of forecasting was also significantly improved by 15% and 13% with hybrid GARCH-LSTM in comparison to single LSTMs. Furthermore, the results reveal that hybrid GARCH-LSTM fully exploits the heteroscedasticity element, which is not captured by the GARCH model estimation, outperforming GARCH models on their own. This finding from this study confirmed that hybrid GARCH-LSTM models are effective forecasting tools for predicting stock price movements. In addition, the proposed model can assist investors in making informed decisions regarding stock prices since it is capable of closely predicting and imitating the observed pattern and trend of KLSE stock prices.

Identification of Counterfeit Videos using Deep Learning Methodology

The rise of deep learning has ushered in a proliferation of deep fake videos, posing significant challenges to the credibility of visual content. Our research introduces a groundbreaking approach by merging Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) to enhance the accuracy of deepfake prediction. This unique integration, which has not been previously implemented, significantly boosts the model's capability to discern deepfakes. The synergy of CNNs and RNNs in our methodology represents an advancement, contributing to increased accuracy in detecting synthetic content. We leverage CNNs and RNNs for an efficient solution. First, we employ a Res-Next CNN to extract distinctive features from individual video frames, effectively encoding spatial information. These features are then used in the subsequent phase, where a LSTM-RNN models temporal dynamics within the videodata.The temporal aspect is crucial in differentiating deep fake videos due to subtle inconsistencies over time. The LSTM RNN processes the feature sequence, enabling the model to identify temporal patterns unique to deep fakes. This holistic approach, combining spatial and temporal analysis, enhances the model's ability to detect even highly convincing synthetic content. Our model is trained on a comprehensive dataset with rigorous evaluations, demonstrating competitive performance through standard metrics such as accuracy, precision. Practically, our model offers real- time video analysis, automatically identifying deep fake content and mitigating potential risks. Importantly, our approach is simple and robust, suitable for deployment across diverse scenarios. In summary, our research provides an effective solution to the critical issue of deepfake detection. By synergizing CNNs and LSTM-based RNNs, we offer a practical means to uphold the integrity of visual content in an era where digital information authenticity is paramount. Keywords— Deep Learning, CNN, RNN, Deepfake, LSTM, accuracy, precision, visual content, digital information

Enhancing the exploitation of natural resources for green energy: An application of LSTM-based meta-model for aluminum prices forecasting

Efficient resource allocation for electric car production can be achieved by anticipating aluminum future pricing. For the electric car sector to maintain a steady supply chain, efficient production, and long-term growth, accurate forecasts are essential for policymakers to plan resource policies. The accuracy of ensemble learning models in predicting Aluminum prices was analyzed using Gene Expression programming and long short-term memory (LSTM) method. The analysis was based on monthly frequency spot settlement price data of Aluminum, copper, silver, and crude oil prices, as well as currency inflation rates of USA, China, and Peru. The study period spanned from January 1, 1994 to December 31, 2022, and also included the gross domestic product (GDP) of USA, China, and other metal minerals. Initially, we assess the level of multicollinearity between each chosen input parameter and Aluminum price volatility using Variance inflation factor-based multicollinearity. The primary innovation of this study lies in the creation of a highly precise model that utilizes the LSTM algorithm to estimate monthly Aluminum prices. This model serves as a revolutionary forecasting method. An extensive evaluation of various ensemble learning models reveals that XGBoost is the optimal model for accurately forecasting monthly Aluminum prices. The LSTM-based recurrent neural network model exhibited the lowest error, with a Mean Relative Error (MRE) ranging from 0.001 to 0.008, RMSE ranging from 54.527 to 136.044, and R2 ranging from 0.95 to 0.983. The findings demonstrated that the LSTM-based super learner models had superior performance compared to the CatBoost, LightBoost, Random Forest, XGBoost, and AdaBoost models, as evidenced by higher values of the determination coefficient (R2), value account for (VAF), and other 10 error analysis evaluators. Therefore, the utilization of an LSTM-based super learner model can serve as a dependable approach for forecasting future monthly Aluminum prices. In addition, this study introduces a logical mathematical model utilizing gene expression programming (GEP) to anticipate future Aluminum prices, which can be utilized by other researchers. Thus, we conclude that both the GEP and ensemble learning models, particularly the LSTM-based Super learner model, are appropriate for precise metal price predictions, which can benefit policymakers and assist in resource policy planning.

Recurrent Neural Networks for Image Captioning: A Case Study with LSTM

This research investigates the viability of Long Short-Term Memory (LSTM) systems, a subtype of Recurrent Neural Networks (RNNs), for picture captioning. Leveraging the MS COCO dataset, the study compares the execution of LSTM-based RNNs with Vanilla RNN, Gated Recurrent Unit (GRU), consideration components, and transformer-based models. Experimental comes about to illustrate that the LSTM-based RNN shows competitive execution, accomplishing a BLEU-4 score of 0.72, a METEOR score of 0.68, and a CIDEr score of 2.1. The comparative investigation uncovers its prevalence over Vanilla RNN and GRU, highlighting its capability to capture long-range conditions inside successive picture information. Moreover, the study investigates the effect of consideration instruments and transformer designs, exhibiting their potential improvements in the context-aware caption era. The transformer-based show outflanks all other models, accomplishing a BLEU-4 score of 0.78, a METEOR score of 0.72, and a CIDEr score of 2.5. The findings give important bits of knowledge toward the creating scene of picture captioning strategy, which makes LSTM-based RNNs solid and productive approaches for capturing worldly groupings in visual substance. In achieving these, the study provides a framework for future developments in hybrid models and manufacturing processes that push boundaries of smart image perception and understanding

Enhancing Electrical Power Demand Prediction Using LSTM-Based Deep Learning Models for Local Energy Communities

The pursuit of accurate electrical power demand forecasting has led to the application of deep learning algorithms, notably demonstrating promising outcomes despite the prerequisite of substantial training data. This study pioneers a new learning paradigm, employing sophisticated deep learning models specifically, Long Short-Term Memory networks and recurrent neural networks (RNNs). Leveraging historical load data, temperature, wind speed, and day-ahead predicted spot prices, this approach follows a structured flow involving data preprocessing, sequence generation, model training, and subsequent prediction of future load demand using LSTM-based RNN variants. The study’s paramount findings underscore the substantial advancement achieved by this proposed methodology over prevailing techniques. The method significantly improves prediction accuracy by over 11%, demonstrating the efficacy of deep learning models and a significant leap forward in forecasting precision. Beyond its superior predictive capabilities, this novel strategy serves as a catalyst for enhancing energy distribution management in local energy communities. Its effectiveness lies not only in its precision but also in enabling the optimization and cost-effective control of energy distribution, vital for sustainable energy management in these communities. Ultimately, this pioneering approach presents a robust solution poised to revolutionize the landscape of electrical power demand forecasting and its practical application in local energy systems.

Continuous monitoring of tissue oxygen metabolism based on multi-wavelength diffuse correlation spectroscopy using LSTM-based RNN model

Growing incidences of cardiovascular diseases have driven demand for continuous tissue oxygen metabolism monitoring. Diffuse optical spectroscopy (DOS) or near-infrared spectroscopy (NIRS) is typically employed to measure oxygenated and de-oxygenated hemoglobin, while diffuse correlation spectroscopy (DCS) provides a direct measure of tissue blood flow noninvasively. Traditionally, combined DCS and DOS/NIRS approaches are utilized for monitoring oxygen metabolism, which needs two different numerical models to calculate blood flow and oxygen saturation. In this paper, we evaluate the utility of multi-wavelength diffuse correlation spectroscopy (mwDCS) using a proposed deep learning approach, i.e., long short-term memory (LSTM) based recurrent neural network (RNN). Arterial occlusion experiments were performed in vivo to generate the data from 6 healthy individuals. Further, based on the dataset acquired by mwDCS, a LSTM-based RNN model was developed to directly assess changes of blood flow and oxygen saturation. Compared to conventional methods, the in vivo experiments in this study demonstrated a more portable system with promising results. Pearsons correlation coefficients of the proposed model reached 0.82 and 0.64 for the changes of blood flow and oxygen saturation, respectively. Our results shows that mwDCS using LSTM-based RNN model provides an alternative method for continuous tissue oxygen metabolism monitoring.

LSTM-based recurrent neural network provides effective short term flu forecasting

BackgroundInfluenza virus is responsible for a yearly epidemic in much of the world. To better predict short-term, seasonal variations in flu infection rates and possible mechanisms of yearly infection variation, we trained a Long Short-Term Memory (LSTM)-based deep neural network on historical Influenza-Like-Illness (ILI), climate, and population data.MethodsData were collected from the Centers for Disease Control and Prevention (CDC), the National Center for Environmental Information (NCEI), and the United States Census Bureau. The model was initially built in Python using the Keras API and tuned manually. We explored the roles of temperature, precipitation, local wind speed, population size, vaccination rate, and vaccination efficacy. The model was validated using K-fold cross validation as well as forward chaining cross validation and compared to several standard algorithms. Finally, simulation data was generated in R and used for further exploration of the model.ResultsWe found that temperature is the strongest predictor of ILI rates, but also found that precipitation increased the predictive power of the network. Additionally, the proposed model achieved a +1 week prediction mean absolute error (MAE) of 0.1973. This is less than half of the MAE achieved by the next best performing algorithm. Additionally, the model accurately predicted simulation data. To test the role of temperature in the network, we phase-shifted temperature in time and found a predictable reduction in prediction accuracy.ConclusionsThe results of this study suggest that short term flu forecasting may be effectively accomplished using architectures traditionally reserved for time series analysis. The proposed LSTM-based model was able to outperform comparison models at the +1 week time point. Additionally, this model provided insight into the week-to-week effects of climatic and biotic factors and revealed potential patterns in data series. Specifically, we found that temperature is the strongest predictor of seasonal flu infection rates. This information may prove to be especially important for flu forecasting given the uncertain long-term impact of the SARS-CoV-2 pandemic on seasonal influenza.

Energy-Efficient Machine Learning for IoT Edge Devices: A Federated Learning Approach

In the realm of modern-day IoT (Internet of Things) deployments, the quest for energy-efficient solutions stands paramount for different use cases. As the IoT ecosystem continues to burgeon, there arises an exigency for resource-conscious algorithms that can effectively navigate the intricacies of data routing while minimizing energy consumption levels. This paper illuminates a pivotal advancement in this domain, introducing a novel paradigm for different scenarios. Existing methodologies have long grappled with the challenge of balancing the ever-increasing computational demands of IoT devices with their constrained power resources. Conventional routing strategies often lack the depth required to optimize energy consumption, resulting in inefficiencies that hamper overall system performance. Prior research has primarily focused on rudimentary routing techniques that fall short in addressing the multifaceted demands of contemporary IoT ecosystems. In response to these limitations, this paper propounds a breakthrough model that leverages the power of Deep Dyna Q integrated with a Long Short-Term Memory (LSTM) based Recurrent Neural Network (RNN). This fusion of cutting-edge technologies not only enhances energy efficiency but also ushers in a new era of intelligent data routing within IoT deployments. The chosen methodology is deliberate; Deep Dyna Q harnesses the power of reinforcement learning to dynamically adapt routing decisions, while the LSTM-based RNN augments the model's ability to capture contextual dependencies crucial in IoT scenarios. This synergy engenders superior energy optimization, transforming the way data is routed in IoT edge devices & samples. The results are resoundingly positive, with empirical evidence showcasing the model's prowess. Tested across multiple contextual datasets, our approach exhibits a remarkable 10.5% improvement in energy efficiency, a commendable 8.5% boost in speed, a 3.9% uptick in throughput, a 4.5% surge in packet delivery ratio, and a notable 4.9% enhancement in consistency when compared against existing methods. Such substantial improvements underscore the transformative impact of our work, setting new standards for energy-efficient data routing in IoT deployments.

Design of ear-contactless stethoscope and improvement in the performance of deep learning based on CNN to classify the heart sound.

Cardiac-related disorders are rapidly growing throughout the world. Accurate classification of cardiovascular diseases is an important research topic in healthcare. During COVID-19, auscultating heart sounds was challenging as health workers and doctors wear protective clothing, and direct contact with patients can spread the outbreak. Thus, contactless auscultation of heart sound is necessary. In this paper, a low-cost ear contactless stethoscope is designed where auscultation is done with the help of a bluetooth-enabled micro speaker instead of an earpiece. The PCG recordings are further compared with other standard electronic stethoscopes like Littman 3M. This work is made to improve the performance of deep learning-based classifiers like recurrent neural networks (RNN) and convolutional neural networks (CNN) for different valvular heart problems using tuning of hyperparameters like learning rate of optimizers, dropout rate, and hidden layer. Hyper-parameter tuning is used to optimize the performances of various deep learning models and their learning curves for real-time analysis. The acoustic, time, and frequency domain features are used in this research. The investigation is made on the heart sounds of normal and diseased patients available from the standard data repository to train the software models. The proposed CNN-based inception network model achieved an accuracy of 99.65 ± 0.06% on the test dataset with a sensitivity of 98.8 ± 0.05% and specificity of 98.2 ± 0.19%. The proposed hybrid CNN-RNN architecture attained 91.17 ± 0.03% accuracy on test data after hyperparameter optimization, whereas the LSTM-based RNN model achieved 82.32 ± 0.11% accuracy. Finally, the evaluated results were compared with machine learning algorithms, and the improved CNN-based Inception Net model is the most effective among others.

A novel transfer learning framework for sorghum biomass prediction using UAV-based remote sensing data and genetic markers.

Yield for biofuel crops is measured in terms of biomass, so measurements throughout the growing season are crucial in breeding programs, yet traditionally time- and labor-consuming since they involve destructive sampling. Modern remote sensing platforms, such as unmanned aerial vehicles (UAVs), can carry multiple sensors and collect numerous phenotypic traits with efficient, non-invasive field surveys. However, modeling the complex relationships between the observed phenotypic traits and biomass remains a challenging task, as the ground reference data are very limited for each genotype in the breeding experiment. In this study, a Long Short-Term Memory (LSTM) based Recurrent Neural Network (RNN) model is proposed for sorghum biomass prediction. The architecture is designed to exploit the time series remote sensing and weather data, as well as static genotypic information. As a large number of features have been derived from the remote sensing data, feature importance analysis is conducted to identify and remove redundant features. A strategy to extract representative information from high-dimensional genetic markers is proposed. To enhance generalization and minimize the need for ground reference data, transfer learning strategies are proposed for selecting the most informative training samples from the target domain. Consequently, a pre-trained model can be refined with limited training samples. Field experiments were conducted over a sorghum breeding trial planted in multiple years with more than 600 testcross hybrids. The results show that the proposed LSTM-based RNN model can achieve high accuracies for single year prediction. Further, with the proposed transfer learning strategies, a pre-trained model can be refined with limited training samples from the target domain and predict biomass with an accuracy comparable to that from a trained-from-scratch model for both multiple experiments within a given year and across multiple years.

An automated patient-specific ECG beat classification using LSTM-based recurrent neural networks

Customized electrocardiogram analysis is inevitable in the cardiac functionality assessment of every single individual patient. In this paper, we firmly introduce a novel framework for patient-specific electrocardiogram (ECG) beat classification characterized by the long-short term memory (LSTM) based recurrent neural networks (RNNs). A novel architecture is developed by governing two LSTM models connected in parallel, and each of them operates on the single-channel ECG input. The proposed model customizes the classification process and imbibes the trait of patient-specific analysis by training the model with the first 5-min segment of heartbeats of the respective patient’s ECG record and tests the same for the rest of the ECG record. The proposed framework undergoes five stages viz; pre-processing, QRS complex detection, segmentation, feature extraction, and LSTM-based RNNs for classification. The efficiency of the proposed framework is deeply examined by the support of various metrics viz; accuracy (Acc), sensitivity (Sen), specificity (Spe), positive predictivity (Ppr), F1 score, and G score and the performance of the proposed method is compared to that of some popular state-of-the-art techniques. The proposed framework is exposed to a highly popular and reliable MIT-BIH arrhythmia database for training and testing purposes. The obtained results are found promising and yield superior classification performance in recognizing ventricular ectopic beats (VEB) and supra-ventricular ectopic beats (SVEB). In addition, the proposed framework is recognized as simple, lightweight yet efficient which in turn makes it suitable for continuous monitoring applications in clinical settings for automated ECG analysis to assist the cardiologist.

Deep-Learning-Driven Secure Offline Blockchain in a Cyber-Physical Social System

A cyber-physical social system (CPSS), which integrates the social world with the cyber world and physical world, plays a crucial role in a more intelligent and efficient network. Recent advances in blockchain have significant impacts on securing the transactions in CPSS. Nevertheless, the deployment of blockchain faces fundamental engineering challenges. Specifically, in remote or disaster areas, users of CPSS typically cannot effectively connect to the blockchain system in real time due to the deficiency of Internet connection. In busy urban areas, however, users are reluctant to frequently connect to the blockchain system due to the costly communication connection. Thus, the offline blockchain system, which enables direct secured offline transaction channels among users in CPSS and efficiently synchronizes batch transactions to the global blockchain system, is a promising solution to address these problems. This article presents a novel CPSS architecture that can realize secure transactions for users participating in offline blockchain transactions by making use of an AI-based algorithm. We first introduce three typical offline blockchain transaction environments in CPSS, followed by analysis of the security issues of the proposed architecture. In addition, we present the hashed time locked contract (HTLC) established for the CPSS users participating in an offline blockchain transaction. To detect potential malicious attackers in offline transactions, we introduce an LSTM-based recurrent neural network (RNN) model to detect the malicious transaction witness. Numerical results are presented to validate the effectiveness of this model.

An Efficient Hybrid Model for Acute Myeloid Leukaemia detection using Convolutional Bi-LSTM based Recurrent Neural Network

ABSTRACT The major aim of this research is to identify the AML early from the morphological images. The proposed approach detects AML through four major stages, namely pre-processing, segmentation, feature extraction and classification. The pre-processing stage is employed for image enhancement. The segmentation process is utilized for segmenting the nucleus as well as the cell mask. When the human is affected by AML, the nucleus shape is asymmetrical, and texture is altered. The features achieved by the feature extraction process are given as the input of the classifier. The classification is attained by using the hybrid convolutional bidirectional LSTM-based recurrent neural network network. Although there are several optimisation algorithms proposed to solve different segmentation problems in the literature, to achieve the classified result more optimal, the slime mould optimisation algorithm is utilised. The data sets to detect AML were collected from Munich University Hospital. The experimental investigation is conducted, and from the experimental outcomes, it is examined that the proposed AML detection has obtained a sensitivity of 95.12%, a specificity of 93.72%, an accuracy of 97.95%, a precision of 94.64%, Matthews correlation coefficient of 98.2% and recall of 95.78%. This indicates that the proposed technique outperforms other state-of-art methods.

DEEPFAKE DETECTION USING EYE-BLINKING PATTERN

Deep learning algorithms have become so potent due to increased computing power that it is now relatively easy to produce human-like synthetic videos, sometimes known as & quot; deep fakes. & quot; It is simple to imagine scenarios in which these realistic face switched deep fakes are used to extort individuals, foment political unrest, and stage fake terrorist attacks. This paper provides a deep learning strategy novel for the efficient separation of fraudulent films produced by AI from actual ones. Automatically spotting replacement and recreation deep fakes is possible with our technology. To combat artificial intelligence, we are attempting to deploy artificial intelligence. The framelevel characteristics are extracted by our system using a Res-Next Convolution neural network, and later these features are applied to train an LSTM-based recurrent neural network to determine if those submitted video is being altered in any way or not, i.e. whether it is a deep fake or authentic video. We test our technique on a sizable quantity of balanced and mixed data sets created by combining the different accessible data sets, such as Face-Forensic++[1], Deep fake detection challenge[2], and Celeb-DF[3], in order to simulate real-time events and improve the model's performance on real-time data.

A generalized ground-motion model for consistent mainshock–aftershock intensity measures using successive recurrent neural networks

Several recent studies have investigated the risk posed to structures by earthquake sequences, proposing state-dependent fragility/vulnerability models for assets in damaged conditions. However, a critical component for such efforts, i.e., ground-motion record selection, has received relatively minor consideration. Specifically, utilization of “consistent” mainshock (MS)–aftershock (AS) ground motions is desirable in practical applications. Such consistency in selecting MS–AS sequences requires proper consideration of the correlations between and within the intensity measures of MS and AS ground motions. Most of the studies in this domain utilize spectral accelerations as the considered ground-motion intensity measures and rely on empirical linear correlation models between the intensity measures of MS and AS ground motions for developing, for instance, record selection approaches. This study proposes a generalized ground-motion model (GGMM) to estimate consistent 30 × 1 vectors of intensity measures for mainshocks (denoted as IMMS) and aftershocks (denoted as IMAS) using a framework of successive long-short-term-memory (LSTM) recurrent neural network (RNN). The vectors of IMMS and IMAS consist of geometric means of significant duration (D_{5 - 95,geom}), Arias intensity (I_{a,geom}), cumulative absolute velocity (CAV_{geom}), peak ground velocity (PGV_{geom}), peak ground acceleration (PGA_{geom}) and RotD50 spectral acceleration (S_{a} left( T right)) at 25 periods for both MS and AS ground motions. The proposed RNN-based GGMM is trained on a carefully selected set of ~ 700 crustal and subduction recorded MS–AS sequences. The inputs to the framework include a 5 × 1 vector of source and site parameters for MS and AS recordings. The residuals of the trained LSTM-based RNN are further used to develop empirical covariance structures for IMMS and IMAS. The proposed framework is finally illustrated to select MS–AS ground motions based on IMMS and IMAS using a multi-criteria objective function. The selected MS–AS ground motion sequences are then used to perform non-linear time-history analyses of a case-study two-spanned symmetric bridge structure. The obtained engineering demand parameters are evaluated and critically discussed.

LSTM-Based RNN Framework to Remove Motion Artifacts in Dynamic Multicontrast MR Images with Registration Model

Today, many people under the age of 10 are being examined for brain-related issues, including tumours, without displaying any symptoms. It is not unusual for children to develop brain-related concerns such as tumours and central nervous system disorders, which may affect 15% of the population. Medical experts believe that the irregular eating habits (junk food) and the consumption of pesticide-tainted fruits and vegetables are to blame. The human body is naturally resistant to harmful gears, but only up to a point. If it exceeds the limit, a cell manipulation process is automatically initiated that can remove dangerous inactive tissues from the cell membrane and later grows into tumour blockage in the human body. Thus, the adoption of an advanced computer-based diagnostic system is highly recommended in order to generate visually enhanced images for anomaly identification and infectious tissue segmentation. In most cases, an MR image is chosen since it is easier to distinguish between affected and nonaffected tissue. Conventional convolution neural network (CCNN) mapping and feature extraction are difficult because of the vast volume of data. In addition, it takes a lengthy time for the MRI scanning process to obtain diverse positions for anomaly identification. Aside from the discomfort, the patient may experience motion abnormalities. Recurrent neural network (RNN) classifies tumour regions into several isolated portions much faster and more accurately, so that it can be prevented. To remove motion artefacts from dynamic multicontrast MR images, a novel long short-term memory- (LSTM-) based RNN framework is introduced in this research. With this method, the MR image’s visual quality is improved over CCNN while simultaneously mapping a larger volume and extracting more quiet characteristics than CCNN can. DC-CNN, SMSR-CNN, FMSI-CNN, and DRCA-CNN results are compared. For both low and high signal-to-noise ratios, the suggested LSTM-based RNN framework has gained reasonable feature intelligibility (SNRs). In comparison to previous approaches, it requires less computing and has higher accuracy when it comes to detecting infected portions.

Determining the Rolling Window Size of Deep Neural Network Based Models on Time Series Forecasting

Time series forecasting has always been a significant task in various domains. In this paper, we propose DeepARMA, a LSTM-based recurrent neural network to tackle this problem. DeepARMA is derived from an existing time series forecasting baseline, DeepAR, overcoming two of its weaknesses: (1) rolling window size determination: the way DeepAR determines rolling window size is casual and vulnerable, which may lead to the unnecessary computation and inefficiency of the model;(2) neglect of the noise: pure autoregressive model cannot deal with the condition where data are composed of various kinds of noise, neither do most of time series models including DeepAR. In order to solve these two problems, we first combine a classic information theoretic criterion, AIC, with the network to determine the proper rolling window size. Then, we propose a jointly-learned neural network fusing white Gaussian noise series given by ARIMA models to DeepAR’s input. That is exactly why we name the network ‘DeepARMA’. Our experiments on a real-world dataset demonstrate that our improvement settles those two problems put forward above.

Deep machine learning with Sentinel satellite data to map paddy rice production stages across West Java, Indonesia

Indonesia recently implemented a novel, technology-driven approach for conducting agricultural production surveys, which involves monthly observations at many thousands of strategic locations and automated data logging via a cellular phone application. Data from these comprehensive field surveys offer immense value for advancing remote sensing technology to map crop production across Indonesia, particularly through the development of machine learning approaches to relate survey data with satellite imagery. The objective of this study was to compare different machine learning scenarios for classifying and mapping the temporal progression of paddy rice production stages across West Java, Indonesia using synthetic aperture radar (SAR) and optical remote sensing data from Sentinel-1 and Sentinel-2 satellites. Monthly paddy rice survey data at 21,696 locations across West Java from November 2018 through April 2019 were used for model training and testing. Five classes related to rice production stage or other field conditions were defined, including rice at tillering, heading, and harvest stages, rice fields with little to no vegetation present, and non-rice areas. A recurrent neural network (RNN) with long short term memory (LSTM) nodes provided optimal performance with classification accuracies of 79.6% and 75.9% for model training and testing, respectively, and reduced computational effort. Other approaches that incorporated a convolutional neural network (CNN) either reduced classification accuracy or increased computational effort. Deep machine learning methods (RNN and CNN) generally outperformed other non-deep classifiers, which achieved up to 63.3% accuracy for model testing. Classification accuracies were optimized by inputting two Sentinel-1 channels (VH and VV polarizations) and ten Sentinel-2 channels. Temporal patterns of paddy rice production stages were consistent between the monthly ground-based agricultural survey data and 10-m, satellite-based rice classification maps obtained by applying the LSTM-based RNN across West Java. The results demonstrated the value of combining modern agricultural survey data, satellite remote sensing, and a recurrent neural network to develop multitemporal maps of paddy rice production stages.

Suggestion and invention of recipes using bi-directional LSTMs-based frameworks

Choosing which recipe to eat and which recipe to avoid isn’t that simple for anyone. It takes strenuous efforts and a lot of time for people to calculate the number of calories and P.H level of the dish. In this paper, we propose an ensemble neural network architecture that suggests recipes based on the taste of the person, P.H level and calorie content of the recipes. We also propose a bi-directional LSTMs-based variational autoencoder for generating new recipes. We have ensembled three bi-directional LSTM-based recurrent neural networks which can classify the recipes based on the taste of the person, P.H level of the recipe and calorie content of the recipe. The proposed model also predicts the taste ratings of the recipes for which we proposed a custom loss function which gave better results than the standard loss functions and the model also predicts the calorie content of the recipes. The bi-directional LSTMs-based variational autoencoder after being trained with the recipes which are fit for the person generates new recipes from the existing recipes. After training and testing the recurrent neural networks and the variational autoencoder, we have tested the model with 20 new recipes and got overwhelming results in the experimentation, the variational autoencoders generated a couple of new recipes, which are healthy to the specific person and will be liked by the specific person.

Predicting the Time Until a Vehicle Changes the Lane Using LSTM-Based Recurrent Neural Networks

To plan safe and comfortable trajectories for automated vehicles on highways, accurate predictions of traffic situations are needed. So far, a lot of research effort has been spent on detecting lane change maneuvers rather than on estimating the point in time a lane change actually happens. In practice, however, this temporal information might be even more useful. This paper deals with the development of a system that accurately predicts the time to the next lane change of surrounding vehicles on highways using long short-term memory-based recurrent neural networks. An extensive evaluation based on a large real-world data set shows that our approach is able to make reliable predictions, even in the most challenging situations, with a root mean squared error around 0.7 seconds. Already 3.5 seconds prior to lane changes the predictions become highly accurate, showing a median error of less than 0.25 seconds. In summary, this article forms a fundamental step towards downstreamed highly accurate position predictions.