Accuracy Of Machine Learning Models Research Articles

Machine learning-based prediction of cognitive outcomes in de novo Parkinson’s disease

Cognitive impairment is a debilitating symptom in Parkinson’s disease (PD). We aimed to establish an accurate multivariate machine learning (ML) model to predict cognitive outcome in newly diagnosed PD cases from the Parkinson’s Progression Markers Initiative (PPMI). Annual cognitive assessments over an 8-year time span were used to define two cognitive outcomes of (i) cognitive impairment, and (ii) dementia conversion. Selected baseline variables were organized into three subsets of clinical, biofluid and genetic/epigenetic measures and tested using four different ML algorithms. Irrespective of the ML algorithm used, the models consisting of the clinical variables performed best and showed better prediction of cognitive impairment outcome over dementia conversion. We observed a marginal improvement in the prediction performance when clinical, biofluid, and epigenetic/genetic variables were all included in one model. Several cerebrospinal fluid measures and an epigenetic marker showed high predictive weighting in multiple models when included alongside clinical variables.

Logistic Regression-Guided Identification of Cofactor Specificity-Contributing Residues in Enzyme with Sequence Datasets Partitioned by Catalytic Properties.

Changing the substrate/cofactor specificity of an enzyme requires multiple mutations at spatially adjacent positions around the substrate pocket. However, this is challenging when solely based on crystal structure information because enzymes undergo dynamic conformational changes during the reaction process. Herein, we proposed a method for estimating the contribution of each amino acid residue to substrate specificity by deploying a phylogenetic analysis with logistic regression. Since this method can estimate the candidate amino acids for mutation by ranking, it is readable and can be used in protein engineering. We demonstrated our concept using redox cofactor conversion of the Escherichia coli malic enzyme as a model, which still lacks crystal structure elucidation. The use of logistic regression with amino acid sequences classified by cofactor specificity showed that the NADP+-dependent malic enzyme completely switched cofactor specificity to NAD+ dependence without the need for a practical screening step. The model showed that surrounding residues made a greater contribution to cofactor specificity than those in the interior of the substrate pocket. These residues might be difficult to identify from crystal structure observations. We show that a highly accurate and inferential machine learning model was obtained using amino acid sequences of structurally homologous and functionally distinct enzymes as input data.

Obese, non-eosinophilic asthma: frequent exacerbators in a real-world setting

Objective In the UK, asthma deaths are at their highest level this century. Increased recognition of at-risk patients is needed. This study phenotyped frequent asthma exacerbators and used machine learning to predict frequent exacerbators. Methods Patients admitted to a district general hospital with an asthma exacerbation between 1st March 2018 and 1st March 2020 were included. Patients were organized into two groups: “Infrequent Exacerbators” (1 admission in the previous 12 months) and “Frequent Exacerbators” (≥2 admissions in the previous 12 months). Patient data were retrospectively collected from hospital and primary care records. Machine learning models were used to predict frequent exacerbators. Results 200 patients admitted for asthma exacerbations were randomly selected (73% female; mean age 47.8 years). Peripheral eosinophilia was uncommon in either group (21% vs 19%). More frequent exacerbators were being treated with high-dose ICS than infrequent exacerbators (46.5% vs 23.2%; P < 0.001), and frequent exacerbators used more SABA inhalers (10.9 vs 7.40; P = 0.01) in the year preceding the current admission. BMI was raised in both groups (34.2 vs 30.9). Logistic regression was the most accurate machine learning model for predicting frequent exacerbators (AUC = 0.80). Conclusions Patients admitted for asthma are predominately female, obese and non-eosinophilic. Patients who require multiple admissions per year have poorer asthma control at baseline. Machine learning algorithms can predict frequent exacerbators using clinical data available in primary care. Instead of simply increasing the dose of corticosteroids, multidisciplinary management targeting Th2-low inflammation should be considered for these patients.

Joint modeling strategy for using electronic medical records data to build machine learning models: an example of intracerebral hemorrhage

BackgroundOutliers and class imbalance in medical data could affect the accuracy of machine learning models. For physicians who want to apply predictive models, how to use the data at hand to build a model and what model to choose are very thorny problems. Therefore, it is necessary to consider outliers, imbalanced data, model selection, and parameter tuning when modeling.MethodsThis study used a joint modeling strategy consisting of: outlier detection and removal, data balancing, model fitting and prediction, performance evaluation. We collected medical record data for all ICH patients with admissions in 2017–2019 from Sichuan Province. Clinical and radiological variables were used to construct models to predict mortality outcomes 90 days after discharge. We used stacking ensemble learning to combine logistic regression (LR), random forest (RF), artificial neural network (ANN), support vector machine (SVM), and k-nearest neighbors (KNN) models. Accuracy, sensitivity, specificity, AUC, precision, and F1 score were used to evaluate model performance. Finally, we compared all 84 combinations of the joint modeling strategy, including training set with and without cross-validated committees filter (CVCF), five resampling techniques (random under-sampling (RUS), random over-sampling (ROS), adaptive synthetic sampling (ADASYN), Borderline synthetic minority oversampling technique (Borderline SMOTE), synthetic minority oversampling technique and edited nearest neighbor (SMOTEENN)) and no resampling, seven models (LR, RF, ANN, SVM, KNN, Stacking, AdaBoost).ResultsAmong 4207 patients with ICH, 2909 (69.15%) survived 90 days after discharge, and 1298 (30.85%) died within 90 days after discharge. The performance of all models improved with removing outliers by CVCF except sensitivity. For data balancing processing, the performance of training set without resampling was better than that of training set with resampling in terms of accuracy, specificity, and precision. And the AUC of ROS was the best. For seven models, the average accuracy, specificity, AUC, and precision of RF were the highest. Stacking performed best in F1 score. Among all 84 combinations of joint modeling strategy, eight combinations performed best in terms of accuracy (0.816). For sensitivity, the best performance was SMOTEENN + Stacking (0.662). For specificity, the best performance was CVCF + KNN (0.987). Stacking and AdaBoost had the best performances in AUC (0.756) and F1 score (0.602), respectively. For precision, the best performance was CVCF + SVM (0.938).ConclusionThis study proposed a joint modeling strategy including outlier detection and removal, data balancing, model fitting and prediction, performance evaluation, in order to provide a reference for physicians and researchers who want to build their own models. This study illustrated the importance of outlier detection and removal for machine learning and showed that ensemble learning might be a good modeling strategy. Due to the low imbalanced ratio (IR, the ratio of majority class and minority class) in this study, we did not find any improvement in models with resampling in terms of accuracy, specificity, and precision, while ROS performed best on AUC.

Accelerated Discovery of Proton-Conducting Perovskites through Density Functional Theory and Machine Learning

Hydrogen is an important energy carrier resource in response to limiting greenhouse gas emissions. Proton-conducting perovskite oxide is one of the key materials for highly efficient carbon-neutral hydrogen technologies, such as hydrogen production, CO2 hydrogenation, and ammonia synthesis. Many attempts have been made based on doped perovskites made of well-tested materials, such as BaZrO3, BaCeO3, BaHfO3, BaTiO3, and SrZrO3. However, the resulting perovskites have often suffered stability and conductivity problems. Furthermore, complex phenomena occurring during hydration present challenges for expanding the materials library. Herein, we demonstrate accelerated discovery of proton-conducting perovskites with high conductivity using machine learning (ML) predictions. We constructed consistent training data using density functional theory (DFT) which enable high accuracy of ML model. DFT computations were performed on > 1000 doped perovskite compositions to get their properties of lattice parameters, point defects (e.g., O vacancies, H interstitials), density of states, hydration energy, and proton migration energy. Several ML algorithms including Linear Regression, Bayesian Ridge Regression, Random Forest Regression, Neural networks, and k-Nearest Neighbor were tested for minimum errors and coefficient of determination. The multidimensional relationships between a set of >50 features and conductivity were mapped out using the optimized ML model. We screened a large material space of A-site and B-site doped perovskites to predict potential proton-conducting materials for various energy applications. The outcomes are promising for accelerating the design and applications of proton-conducting perovskite oxides in hydrogen technologies.

Toward reliable machine learning with Congruity: a quality measure based on formal concept analysis

The spreading of machine learning (ML) and deep learning (DL) methods in different and critical application domains, like medicine and healthcare, introduces many opportunities but raises risks and opens ethical issues, mainly attaining to the lack of transparency. This contribution deals with the lack of transparency of ML and DL models focusing on the lack of trust in predictions and decisions generated. In this sense, this paper establishes a measure, namely Congruity, to provide information about the reliability of ML/DL model results. Congruity is defined by the lattice extracted through the formal concept analysis built on the training data. It measures how much the incoming data items are close to the ones used at the training stage of the ML and DL models. The general idea is that the reliability of trained model results is highly correlated with the similarity of input data and the training set. The objective of the paper is to demonstrate the correlation between the Congruity and the well-known Accuracy of the whole ML/DL model. Experimental results reveal that the value of correlation between Congruity and Accuracy of ML model is greater than 80% by varying ML models.

Identifying well-folded de novo proteins in the new era of accurate structure prediction.

Computational de novo protein design tailors proteins for target structures and oligomerisation states with high stability, which allows overcoming many limitations of natural proteins when redesigned for new functions. Despite significant advances in the field over the past decade, it remains challenging to predict sequences that will fold as stable monomers in solution or binders to a particular protein target; thereby requiring substantial experimental resources to identify proteins with the desired properties. To overcome this, here we leveraged the large amount of design data accumulated in the last decade, and the breakthrough in protein structure prediction from last year to investigate on improved ways of selecting promising designs before experimental testing. We collected de novo proteins from previous studies, 518 designed as monomers of different folds and 2112 as binders against the Botulinum neurotoxin, and analysed their structures with AlphaFold2, RoseTTAFold and fragment quality descriptors in combination with other properties related to surface interactions. These features showed high complementarity in rationalizing the experimental results, which allowed us to generate quite accurate machine learning models for predicting well-folded monomers and binders with a small set of descriptors. Cross-validating designs with varied orthogonal computational techniques should guide us for identifying design imperfections, rescuing designs and making more robust design selections before experimental testing.

Multilayer dynamic ensemble model for intensive care unit mortality prediction of neonate patients

Robust and rabid mortality prediction is crucial in intensive care units because it is considered one of the critical steps for treating patients with serious conditions. Combining mortality prediction with the length of stay (LoS) prediction adds another level of importance to these models. No studies in the literature predict such tasks for neonates, especially using time-series data and dynamic ensemble techniques. Dynamic ensembles are novel techniques that dynamically select the base classifiers for each new case. Medically, implementing an accurate machine learning model is insufficient to gain the trust of physicians. The model must be able to justify its decisions. While explainable AI (XAI) techniques can be used to handle this challenge, no studies have been done in this regard for neonate monitoring in the neonatal intensive care unit (NICU). This study utilizes advanced machine learning approaches to predict mortality and LoS through data-driven learning. We propose a multilayer dynamic ensemble-based model to predict mortality as a classification task and LoS as a regression task for neonates admitted to the NICU. The model has been built based on the patient’s time-series data of the first 24 h in the NICU. We utilized a cohort of 3,133 infants from the MIMIC-III real dataset to build and optimize the selected algorithms. It has shown that the dynamic ensemble models achieved better results than other classifiers, and static ensemble regressors achieved better results than classical machine learning regressors. The proposed optimized model is supported by three well-known explainability techniques of SHAP, decision tree visualization, and rule-based system. To provide online assistance to physicians in monitoring and managing neonates in the NICU, we implemented a web-based clinical decision support system based on the most accurate models and selected XAI techniques. The code of the proposed models is publicly available at https://github.com/InfoLab-SKKU/neonateMortalityPrediction.

An Extreme Gradient Boosting approach to estimate the shear strength of FRP reinforced concrete beams

Although fiber-reinforced polymer (FRP) bars have gained much attention recently, their material characteristics, particularly the elastic modulus, are markedly different from conventional reinforcement. This explains why classical calculation models cannot correctly capture the behaviors of FRP-reinforced concrete (RC) beams in structural engineering. As a novel approach, this study aims to construct a machine learning (ML) model for predicting the shear strength (SS) of FRP-RC beams with and without stirrups. An Extreme Gradient Boosting (XGB) model is selected thanks to its effectiveness and robustness in tackling many complicated problems. An experimental database encompassing 453 examples is utilized, with input variables covering the FRP-RC beams’ geometry, the mechanical properties of concrete, and two reinforcement components. The selection of hyperparameters of XGB model is first conducted, followed by a prediction performance assessment using common statistical measures. Additionally, this study conducts feature importance analysis and employs SHAP values technique to discover the inputs versus output relationship, as well as to identify the most significant features affecting the SS of FRP-RC beams. The outcome of this research, which lies in developing a reliable and accurate ML model that can handle both FRP-RC beams with and without stirrups, may favor the application of ML models to various real-world engineering problems.

Toward Generative Adversarial Networks for the Industrial Internet of Things

Machine learning, as a viable way of conducting data analytics, has been successfully applied to a number of areas. Nonetheless, the lack of sufficient data is one critical issue for applying machine learning in Industrial Internet of Things (IIoT) systems. Insufficient data raises could negatively affect the accuracy of machine learning models. To tackle this issue, we design a framework to systematically investigate the impacts of insufficient data on model training. This framework employs the generative adversarial network (GAN) and continuous learning to generate and engage new data in model training, enabling us to study the security risks of introducing new data in the model training process and develop countermeasures to mitigate these risks. To validate the efficacy of our framework, we consider a representative IIoT scenario, in which a variety of industrial components needs to be recognized by convolutional neural networks (CNNs), and design and implement three evaluation scenarios that are based on a real-world IIoT data set. Our experimental results confirm that insufficient data can have a significant impact on the model accuracy, but that new data generated by GAN and continuous learning can greatly improve the model accuracy. Our experimental results also show that the data poisoning threat posed by the GAN can significantly reduce the model accuracy. However, our proposed defensive mechanism is capable of securing the model learning process. We conclude this article by discussing some emerging issues that need to be addressed in future work.

Machine learning-based inverse design for single-phase high entropy alloys

In this work, we develop an inverse design framework to search for single-phase high entropy alloys (HEAs) subjected to specified phase targets and constraints. This framework is based on the fast grid search in the composition–temperature space, enabled by a highly accurate and efficient machine learning model trained by a huge amount of data. Using the framework, we search through the entire quaternary, quinary, and senary alloy systems, formed by Al, Co, Cr, Cu, Fe, Mn, Ni, and Ti, to identify three types of HEAs: (1) the single-phase FCC HEA with the highest Al content; (2) the single-phase FCC HEA with lower equilibrium temperatures; and (3) single-phase BCC HEAs with Al as the principal element. For the first time, we reveal that the highest Al content in single-phase FCC HEAs is 0.15 in mole fraction, which is higher than the Al contents in all reported single-phase FCC HEAs. The identified HEAs for the quaternary, quinary, and senary groups are Al0.15Co0.34Cr0.16Ni0.35, Al0.15Co0.35Cr0.1Fe0.05Ni0.35, and Al0.15Co0.36Cr0.06Fe0.06Mn0.01Ni0.36, respectively. All the designed HEAs are verified by the equilibrium calculations with Thermo-Calc software and the TCHEA3 database. We further conduct Scheil–Gulliver calculations and experimental fabrications and characterizations for the designed HEAs, to verify the formation of the targeted phases at non-equilibrium conditions. This work demonstrates a viable approach to design HEAs with specified phase targets and constraints.

Inconsistency-Based Multi-Task Cooperative Learning for Emotion Recognition

Emotion recognition is an important part of affective computing. Human emotions can be described categorically or dimensionally. Accurate machine learning models for emotion classification and estimation usually depend on a large amount of annotated data. However, label acquisition in emotion recognition is costly: obtaining the ground-truth labels of an emotional sample usually requires multiple annotators’ assessments, which is expensive and time-consuming. To reduce the labeling effort in multi-task emotions recognition, the paper proposes an inconsistency measure that can indicate the difference between the labels estimated from the feature space and the label distribution of labeled dataset. Using the inconsistency as an indicator of sample informativeness, we further propose an inconsistency-based multi-task cooperative learning framework that integrates multi-task active learning and self-training semi-supervised learning. Experiments in two multi-task emotion recognition scenarios, multi-dimensional emotion estimation and simultaneous emotion classification and estimation, were conducted under this framework. The results demonstrated that the proposed multi-task active learning framework outperformed several single-task and multi-task active learning approaches.

Association between Self-Reported Prior Night's Sleep and Single-Task Gait in Healthy, Young Adults: A Study Using Machine Learning.

Failure to obtain the recommended 7–9 h of sleep has been associated with injuries in youth and adults. However, most research on the influence of prior night’s sleep and gait has been conducted on older adults and clinical populations. Therefore, the objective of this study was to identify individuals who experience partial sleep deprivation and/or sleep extension the prior night using single task gait. Participants (n = 123, age 24.3 ± 4.0 years; 65% female) agreed to participate in this study. Self-reported sleep duration of the night prior to testing was collected. Gait data was collected with inertial sensors during a 2 min walk test. Group differences (<7 h and >9 h, poor sleepers; 7–9 h, good sleepers) in gait characteristics were assessed using machine learning and a post-hoc ANCOVA. Results indicated a correlation (r = 0.79) between gait parameters and prior night’s sleep. The most accurate machine learning model was a Random Forest Classifier using the top 9 features, which had a mean accuracy of 65.03%. Our findings suggest that good sleepers had more asymmetrical gait patterns and were better at maintaining gait speed than poor sleepers. Further research with larger subject sizes is needed to develop more accurate machine learning models to identify prior night’s sleep using single-task gait.

Energy Forecasting in a Public Building: A Benchmarking Analysis on Long Short-Term Memory (LSTM), Support Vector Regression (SVR), and Extreme Gradient Boosting (XGBoost) Networks

A primary energy consumption and CO2 emission source stems from buildings and infrastructures due to rapid urbanisation and social development. An accurate method to forecast energy consumption in a building is thus critically needed to enable successful management of adaptive energy consumption and ease the level of CO2 emission. However, energy forecasting for buildings, especially residential buildings, has several challenges, such as significant variations in energy usage patterns due to unpredicted demands of the residences and some intricate factors, which can randomly affect the patterns. Traditional forecasting approaches require a tremendous number of inputs needed for building physic models and variations often exist between as-built and as-designed buildings in reality. Most recent studies have adopted only ambient weather conditions, building components, and the occupant’s behaviours. As a result, in order to take into account the complexity of factors that can affect the building energy model development and its computation, we develop advanced machine learning models driven by the inherent electricity consumption pattern associated with the day and time. In this study, we demonstrate benchmarking results derived from three different machine learning algorithms, namely SVR, XGBoost, and LSTM, trained by using 1-year datasets with sub-hourly (30 min) temporal granularity to determine the outperformed predictor. Ultimately, the machine learning model robustness and performance on a basis of the coefficient of variation (CV) obtained by the SVR is benchmarked across XGBoost and LSTM trained by the same datasets containing attributes related to the building type, data size, and temporal granularity. The insight stemming from this study indicates that the suitable choice of the machine learning models for building energy forecasts largely depends on the natural characteristics of building energy data. Hyperparameter tuning or mathematical modification within an algorithm may not be sufficient to attain the most accurate machine learning model for building energy forecast.

Machine Learning Algorithms for understanding the determinants of under-five Mortality

BackgroundUnder-five mortality is a matter of serious concern for child health as well as the social development of any country. The paper aimed to find the accuracy of machine learning models in predicting under-five mortality and identify the most significant factors associated with under-five mortality.MethodThe data was taken from the National Family Health Survey (NFHS-IV) of Uttar Pradesh. First, we used multivariate logistic regression due to its capability for predicting the important factors, then we used machine learning techniques such as decision tree, random forest, Naïve Bayes, K- nearest neighbor (KNN), logistic regression, support vector machine (SVM), neural network, and ridge classifier. Each model’s accuracy was checked by a confusion matrix, accuracy, precision, recall, F1 score, Cohen’s Kappa, and area under the receiver operating characteristics curve (AUROC). Information gain rank was used to find the important factors for under-five mortality. Data analysis was performed using, STATA-16.0, Python 3.3, and IBM SPSS Statistics for Windows, Version 27.0 software.ResultBy applying the machine learning models, results showed that the neural network model was the best predictive model for under-five mortality when compared with other predictive models, with model accuracy of (95.29% to 95.96%), recall (71.51% to 81.03%), precision (36.64% to 51.83%), F1 score (50.46% to 62.68%), Cohen’s Kappa value (0.48 to 0.60), AUROC range (93.51% to 96.22%) and precision-recall curve range (99.52% to 99.73%). The neural network was the most efficient model, but logistic regression also shows well for predicting under-five mortality with accuracy (94% to 95%)., AUROC range (93.4% to 94.8%), and precision-recall curve (99.5% to 99.6%). The number of living children, survival time, wealth index, child size at birth, birth in the last five years, the total number of children ever born, mother’s education level, and birth order were identified as important factors influencing under-five mortality.ConclusionThe neural network model was a better predictive model compared to other machine learning models in predicting under-five mortality, but logistic regression analysis also shows good results. These models may be helpful for the analysis of high-dimensional data for health research.

Daily Local-Level Estimates of Ambient Wildfire Smoke PM2.5 for the Contiguous US.

Smoke from wildfires is a growing health risk across the US. Understanding the spatial and temporal patterns of such exposure and its population health impacts requires separating smoke-driven pollutants from non-smoke pollutants and a long time series to quantify patterns and measure health impacts. We develop a parsimonious and accurate machine learning model of daily wildfire-driven PM2.5 concentrations using a combination of ground, satellite, and reanalysis data sources that are easy to update. We apply our model across the contiguous US from 2006 to 2020, generating daily estimates of smoke PM2.5 over a 10 km-by-10 km grid and use these data to characterize levels and trends in smoke PM2.5. Smoke contributions to daily PM2.5 concentrations have increased by up to 5 μg/m3 in the Western US over the last decade, reversing decades of policy-driven improvements in overall air quality, with concentrations growing fastest for higher income populations and predominantly Hispanic populations. The number of people in locations with at least 1 day of smoke PM2.5 above 100 μg/m3 per year has increased 27-fold over the last decade, including nearly 25 million people in 2020 alone. Our data set can bolster efforts to comprehensively understand the drivers and societal impacts of trends and extremes in wildfire smoke.

OES-Fed: a federated learning framework in vehicular network based on noise data filtering.

The Internet of Vehicles (IoV) is an interactive network providing intelligent traffic management, intelligent dynamic information service, and intelligent vehicle control to running vehicles. One of the main problems in the IoV is the reluctance of vehicles to share local data resulting in the cloud server not being able to acquire a sufficient amount of data to build accurate machine learning (ML) models. In addition, communication efficiency and ML model accuracy in the IoV are affected by noise data caused by violent shaking and obscuration of in-vehicle cameras. Therefore we propose a new Outlier Detection and Exponential Smoothing federated learning (OES-Fed) framework to overcome these problems. More specifically, we filter the noise data of the local ML model in the IoV from the current perspective and historical perspective. The noise data filtering is implemented by combining data outlier, K-means, Kalman filter and exponential smoothing algorithms. The experimental results of the three datasets show that the OES-Fed framework proposed in this article achieved higher accuracy, lower loss, and better area under the curve (AUC). The OES-Fed framework we propose can better filter noise data, providing an important domain reference for starting field of federated learning in the IoV.

Prediction of Nitrogen Dosage in 'Alicante Bouschet' Vineyards with Machine Learning Models.

Vineyard soils normally do not provide the amount of nitrogen (N) necessary for red wine production. Traditionally, the N concentration in leaves guides the N fertilization of vineyards to reach high grape yields and chemical composition under the ceteris paribus assumption. Moreover, the carryover effects of nutrients and carbohydrates stored by perennials such as grapevines are neglected. Where a well-documented database is assembled, machine learning (ML) methods can account for key site-specific features and carryover effects, impacting the performance of grapevines. The aim of this study was to predict, using ML tools, N management from local features to reach high berry yield and quality in ‘Alicante Bouschet’ vineyards. The 5-year (2015–2019) fertilizer trial comprised six N doses (0–20–40–60–80–100 kg N ha−1) and three regimes of irrigation. Model features included N dosage, climatic indices, foliar N application, and stem diameter of the preceding season, all of which were indices of the carryover effects. Accuracy of ML models was the highest with a yield cutoff of 14 t ha−1 and a total anthocyanin content (TAC) of 3900 mg L−1. Regression models were more accurate for total soluble solids (TSS), total titratable acidity (TTA), pH, TAC, and total phenolic content (TPC) in the marketable grape yield. The tissue N ranges differed between high marketable yield and TAC, indicating a trade-off about 24 g N kg−1 in the diagnostic leaf. The N dosage predicted varied from 0 to 40 kg N ha−1 depending on target variable, this was calculated from local features and carryover effects but excluded climatic indices. The dataset can increase in size and diversity with the collaboration of growers, which can help to cross over the numerous combinations of features found in vineyards. This research contributes to the rational use of N fertilizers, but with the guarantee that obtaining high productivity must be with adequate composition.

ZERO INFRASTRUCTURE GEOLOCATION OF NEARBY FIRST RESPONDERS ON RO-RO VESSELS

The early suppression of fires on ro-ro vessels requires rapid fire identification as a fire of medium growth exponentially reaches 50kW after only 1 minute. Fire patrol members (e.g., able seamen) are asked to act as first responders in such fire incident cases. They do however lack the necessary digital technology for immediate localization, verification and coordination with the bridge and other first responders. Indoor localization requires dense referencing systems (such as Wi-Fi, UWB, Bluetooth antennas), but these technologies require expensive installations and maintenance. Also, Satellite-based indoor localization is obstructed by the bulky steel structures of vessels, so this doesn’t work either. Within the LASH FIRE project, an H2020 funded project (Grant Agreement #814975) in which this publication is framed, research has been carried out to develop a ground-breaking localization technology that requires zero infrastructure using computer vision on commodity smartphone devices attached to the gear of first responders. The developed solution comprises of three steps: (i) Training, where vessel owners supply video recordings that are processed on a deep learning data center to produce an accurate computer vision machine learning model; (ii) Logging, where a mobile app allows referencing non-movable objects to the (x,y,deck) coordinates of a vessel; and (iii) Localization, where first responders localize on a digital map. Additionally, in case a sparse communication network is available, first responders can share their location, emergency messages and heat scan images with nearby first responders and the bridge. Our proposed technology is shown to be 80% and 90% accurate for localization and tracking scenarios, respectively, in a study we carried out with video traces from a real ro-ro vessel. The overall developed Smart Alert System (SMAS), streamlines the lengthy fire verification, coordination, and reaction process in the early stages of a fire, improving fire safety.

Safe semi-supervised learning using a bayesian neural network

Semi-supervised learning attempts to use a large set of unlabeled data to increase the prediction accuracy of machine learning models when the amount of labeled data is limited. However, in realistic cases, unlabeled data may worsen performance because they contain out-of-distribution (OOD) data that differ from the labeled data. To address this issue, safe semi-supervised deep learning has recently been presented. This study suggests a new safe semi-supervised algorithm that uses an uncertainty-aware Bayesian neural network. Our proposed method, safe uncertainty-based consistency training (SafeUC), uses Bayesian uncertainty to minimize the harmful effects caused by unlabeled OOD examples. The proposed method improves the model’s generalization performance by regularizing the network for consistency against uncertain noise. Moreover, to avoid uncertain prediction results, the proposed method includes a practical inference tip based on a well-calibrated uncertainty. The effectiveness of the proposed method is demonstrated in the experimental results on CIFAR-10 and SVHN by showing that it achieved state-of-the-art performance for all semi-supervised learning tasks with OOD data presence rates.