Random Training Samples Research Articles

Improving radial lens distortion correction with multi-task learning

With computer vision and machine learning, sports image analysis can enhance viewer engagement and contribute to the overall understanding of sports events. However, beneath the apparent simplicity of sports images lies a complex challenge of radial distortion, which can impede their accurate interpretation. The need for high precision and real-time performance is paramount. We present a new regression-based method for radial distortion correction that improves the accuracy of distortion model coefficient prediction by introducing a secondary learning task that compares the distortion level of two random training samples. The secondary task requires no additional annotation, and because it is also related to radial distortion, it encourages the common feature extractor component to learn more general and robust features while preventing overfitting and improving the efficiency of training data. We have evaluated our proposed method using two public datasets and compared it against five other methods. Our method surpassed them all, both in the accuracy of the radial distortion correction and speed as well. You can find the source code and trained models at https://vgg.fiit.stuba.sk/improving-radial.

Enhancing vegetation formation classification: Integrating coarse-scale traditional mapping knowledge and advanced machine learning

Mapping vegetation formation types in large areas is crucial for ecological and environmental studies. However, this is still challenging to distinguish similar vegetation formation types using existing predictive vegetation mapping methods, based on commonly used environmental variables and remote sensing spectral data, especially when there are not enough training samples. To solve this issue, we proposed a predictive vegetation mapping method by integrating an advanced machine learning algorithm and knowledge in an early coarse-scale vegetation map (VMK). First, we implemented classification using the random forest algorithm by integrating the early vegetation map as an auxiliary feature (VMF). Then, we determined the rationality of classified vegetation types and distinguished the confusing types, respectively, based on the knowledge of the spatial distributions and hierarchies of vegetation. Finally, we replaced each recognized unreasonable vegetation type with its corresponding reasonable vegetation type. We implemented the new method in upstream of the Yellow River based on GaoFen-1 satellite images and other environmental variables (i.e., topographical and climate variables). Results showed that the overall accuracy using the VMK method ranged from 67.7 % to 76.8 %, which was 10.9 % to 13.4 % and 3.2 % to 6.6 %, respectively, higher than that of the method without the early vegetation map (NVM) and the VMF method, based on cross-validation with 20 % to 60 % random training samples. The spatial details of the vegetation map using the VMK method were also more reasonable compared to the NVM and VMF methods. These results indicated that the VMK method can distinctly improve the mapping accuracy at the vegetation formation level by integrating knowledge of existing vegetation maps. The proposed method can largely reduce the requirements on the number of field samples, which is especially important for alpine mountains and arctic region, where collecting training samples is more difficult due to the harsh natural environment.

UDL: a cloud task scheduling framework based on multiple deep neural networks

Cloud task scheduling and resource allocation (TSRA) constitute a core issue in cloud computing. Batch submission is a common user task deployment mode in cloud computing systems. In this mode, it has been a challenge for cloud systems to balance the quality of user service and the revenue of cloud service provider (CSP). To this end, with multi-objective optimization (MOO) of minimizing task latency and energy consumption, we propose a cloud TSRA framework based on deep learning (DL). The system solves the TSRA problems of multiple task queues and virtual machine (VM) clusters by uniting multiple deep neural networks (DNNs) as task scheduler of cloud system. The DNNs are divided into exploration part and exploitation part. At each scheduling time step, the model saves the best outputs of all scheduling policies from each DNN to the experienced sample memory pool (SMP), and periodically selects random training samples from SMP to train each DNN of exploitation part. We designed a united deep learning (UDL) algorithm based on this framework. Experimental results show that the UDL algorithm can effectively solve the MOO problem of TSRA for cloud tasks, and performs better than benchmark algorithms such as heterogeneous distributed deep learning (HDDL) in terms of task scheduling performance.

Wind speed retrieval using GNSS-R technique with geographic partitioning

In this paper, the effect of geographical location on Cyclone Global Navigation Satellite System (CYGNSS) observables is demonstrated for the first time. It is found that the observables corresponding to the same wind speed vary with geographic location regularly. Although latitude and longitude information is included in the conventional method, it cannot effectively reduce the errors caused by geographic differences due to the non-monotonic changes of observables with respect to latitude and longitude. Thus, an improved method for Global Navigation Satellite System Reflectometry (GNSS-R) wind speed retrieval that takes geographical differences into account is proposed. The sea surface is divided into different areas for independent wind speed retrieval, and the training set is resampled by considering high wind speed. To balance between the retrieval accuracies of high and low wind speeds, the results with the random training samples and the resampling samples are fused. Compared with the conventional method, in the range of 0–20 m/s, the improved method reduces the Root Mean Square Error (RMSE) of retrieved wind speeds from 1.52 to 1.34 m/s, and enhances the correlation coefficient from 0.86 to 0.90; while in the range of 20–30 m/s, the RMSE decreases from 8.07 to 4.06 m/s, and the correlation coefficient increases from 0.04 to 0.45. Interestingly, the SNR observations are moderately correlated with marine gravities, showing correlation coefficients of 0.5–0.6, which may provide a useful reference for marine gravity retrieval using GNSS-R in the future.

Assessing the effectiveness of alternative landslide partitioning in machine learning methods for landslide prediction in the complex Himalayan terrain

Several devastating landslides have occurred in the NW Himalayas, which has prompted several researchers to strive for improvement in landslide susceptibility modelling (LSM) methodologies. This research analyzes the effectiveness of alternative landslide partitioning techniques on LSM in the landslide-prone district, Muzaffarabad, Pakistan. We developed a landslide inventory of 961 landslides and then traditionally divided it into training (672; 70%) and testing (289; 30%) samples. These training samples (672) are processed by the Average Nearest Neighbour Index (ANNI) method to estimate the spatial pattern of landslides in nature. The results provide an ANNI ratio of 0.672 confirming that the landslides distribution pattern is cluster in the complex Himalayan terrain of Muzaffarabad. Among 672, the majority of landslides (529; 79%) depict cluster behaviour, while 189 landslides (21%) depict random behaviour. To evaluate the effectiveness of landslide cluster samples in prediction, five machine learning algorithms (MLAs), that is, K-Nearest Neighbour (KNN), Naïve Bayes (NB), Random Forest (RF), Extreme Gradient Boosting (XGBoost) and Logistic Regression (LR) using proposed cluster (529) and traditional (672) random training samples along with 17 geo-environmental factors are executed. However, testing samples (289; 30%) separated at the initial stage remained the same to check the model’s effectiveness. The areas under the curve (AUC-ROC), sensitivity, specificity, Kappa index and accuracy (ACC) have been used to evaluate the MLA’s performances. An alternative partitioning technique (cluster) shows the highest predictive power with AUC-ROC values ranging from 0.96 to 0.86, Kappa index ranges from 0.76 to 0.60 and ACC ranges from 0.90 to 0.83. Conversely, the random partitioning approach performs less well with AUC-ROC values ranging from 0.95 to 0.83, Kappa index ranges from 0.70 to 0.49 and ACC ranges from 0.87 to 0.80. In comparison, the RF cluster sampling-based model outperforms the other models and their counterparts. The RF model achieved the highest accuracy (0.902), highest AUC (0.962) and highest Kappa index (0.755) followed by XGboost having ACC (0.885), AUC (0.95) and Kappa index (0.724) employing proposed cluster training samples. However, traditional random training samples yield comparatively low ACC of RF (0.868) and XGboost (0.862). These results confirm that cluster training sampling performs well in obtaining reliable and precise LSMs for complex Himalayan terrain. Although random landslide partitioning for training datasets is seldom utilized in LSM, this study highlights that cluster partitioning for landslide training datasets might be a realistic and reliable approach.

Study on the semi-supervised learning-based patient similarity from heterogeneous electronic medical records

BackgroundA new learning-based patient similarity measurement was proposed to measure patients’ similarity for heterogeneous electronic medical records (EMRs) data.MethodsWe first calculated feature-level similarities according to the features’ attributes. A domain expert provided patient similarity scores of 30 randomly selected patients. These similarity scores and feature-level similarities for 30 patients comprised the labeled sample set, which was used for the semi-supervised learning algorithm to learn the patient-level similarities for all patients. Then we used the k-nearest neighbor (kNN) classifier to predict four liver conditions. The predictive performances were compared in four different situations. We also compared the performances between personalized kNN models and other machine learning models. We assessed the predictive performances by the area under the receiver operating characteristic curve (AUC), F1-score, and cross-entropy (CE) loss.ResultsAs the size of the random training samples increased, the kNN models using the learned patient similarity to select near neighbors consistently outperformed those using the Euclidean distance to select near neighbors (all P values < 0.001). The kNN models using the learned patient similarity to identify the top k nearest neighbors from the random training samples also had a higher best-performance (AUC: 0.95 vs. 0.89, F1-score: 0.84 vs. 0.67, and CE loss: 1.22 vs. 1.82) than those using the Euclidean distance. As the size of the similar training samples increased, which composed the most similar samples determined by the learned patient similarity, the performance of kNN models using the simple Euclidean distance to select the near neighbors degraded gradually. When exchanging the role of the Euclidean distance, and the learned patient similarity in selecting the near neighbors and similar training samples, the performance of the kNN models gradually increased. These two kinds of kNN models had the same best-performance of AUC 0.95, F1-score 0.84, and CE loss 1.22. Among the four reference models, the highest AUC and F1-score were 0.94 and 0.80, separately, which were both lower than those for the simple and similarity-based kNN models.ConclusionsThis learning-based method opened an opportunity for similarity measurement based on heterogeneous EMR data and supported the secondary use of EMR data.

Guided MDNet tracker with guided samples

Visual tracking is the process of estimating the position of an object in a video sequence and plays a very important role in the field of autonomous video processing. Recent work renders that the trackers developed using deep learning techniques such as the convolutional neural network (CNN) exhibits outstanding performances in terms of accuracy and robustness as compared to other state-of-the-art trackers. Multi-domain convolutional neural network (MDNet) is a deep tracker which uses the CNN for estimating the target in each frame of the video sequence. The majority of the tracking challenges could be very easily handled by the MDNet tracker due to its offline training and online tracking features. The offline training stage helps in capturing the target representations into the shared layers of the CNN, while the online tracking uses a large number of probable random samples of bboxes (bounding boxes) around the previous target for estimating the target in the current frame. Once the target is estimated, a process of fine-tuning is performed which will update the weights of shared layers of CNN. The large number of random samples used for target estimation and the huge number of random training samples generated for the fine-tuning during the online stage makes tracking by MDNet computationally complex and slow. The major contribution of this paper is to suggest using guided samples to the input of the CNN rather than random samples. Moreover, it generates a lesser number of highly efficient training samples for the fine-tuning which helps in decreasing the computational complexity of the tracker by half without much compromise on the performance and thus improves the speed of the tracking process. An extensive evaluation has been performed on the proposed Guided MDNet with different datasets like ALOV300++, OTB and VOT, and its performances are measured in terms of metrics like F-score, one-pass evaluation, robustness and accuracy.

An Iterative Random Training Sample Selection Approach to Constrained Energy Minimization for Hyperspectral Image Classification

Iterative constrained energy minimization (ICEM) has shown success in classification. However, a drawback suffered from ICEM is its requirement of complete ground truth to calculate class means. This letter develops an iterative selection of training samples to extend ICEM with two versions: iterative fixed training sampling constrained energy minimization (CEM) (IFTS-CEM) which uses a fixed training sample set throughout the entire iterative process and iterative random training sampling CEM (IRTS-CEM) which uses a random training sampling (RTS) at each iteration. The experimental results demonstrate that IRTS-CEM performs better than IFTS-CEM and also comparable to ICEM.

Annual large-scale urban land mapping based on Landsat time series in Google Earth Engine and OpenStreetMap data: A case study in the middle Yangtze River basin

Long time series (e.g. 30 years) urban land observations from remote sensing images are important for urban growth modeling as well as for the goal of sustainable urban development. However, updates to regional and even global maps are infrequent due to the cost and difficulty of collecting representative training data and the requirement for high-performance computation in processing large amounts of images. In this study, a semi-automatic large-scale and long time series (LSLTS) urban land mapping framework is demonstrated by integrating the crowdsourced OpenStreetMap (OSM) data with free Landsat images (~13,218 scenes) to generate annual urban land maps in the 317,000 km2 middle Yangtze River basin (MYRB) from 1987 to 2017 facilitated by Google Earth Engine (GEE). Random training samples for latest year were generated based on the updated OSM land use data after a manual topological conflict processing and uploaded to GEE for automatic image classification. For each historical year, training samples were obtained with a proposed transferring schema by which only the unchanged were selected through a change detection analysis. The annual spectral indices and texture feature maps acquired from the surface reflectance dataset were also added to the original bands. Finally, the classified maps were downloaded from GEE and a spatial-temporal consistency checking was further performed. Based on independent samples, the overall accuracies and kappa coefficients of all years ranged from 98% to 99% and 0.65 to 0.85, respectively. Our product when compared with current 30 m land-cover products showed similar accuracies but more spatial details. The characteristics of pattern, traces, and hotspots of urban expansion were further explored. This study provides a more convenient procedure for LSLTS urban land mapping especially for areas where large-scale field sample-collection is difficult and little historical crowdsourced datasets are available. The resultant dataset is expected to provide consistent details about the spatial distribution of urban land in MYRB. We highlight the potential use of this proposed framework to be applied and validated to other parts of the world to help better understand and quantify various aspects of urban-related problems.

Classification based on prototypes with spheres of influence

We present a family of binary classifiers and analyse their performance. Each classifier is determined by a set of ‘prototypes’, with given labels. The classification of a given point is determined through the sign of a discriminant function. For each prototype, its sphere of influence is the largest sphere centred on it that contains no prototypes of opposite label, and, given a point to be classified, there is a contribution to the discriminant function at that point from precisely those prototypes whose spheres of influence contain the point. This contribution is positive from positive prototypes and negative from negative prototypes. These contributions are larger in absolute value the closer the point is (relative to the sphere's radius) to the prototype. We quantify the generalization error of such classifiers in a standard probabilistic learning model which involves the values of the discriminant function on the points of a random training sample.

Social media research: The application of supervised machine learning in organizational communication research.

Despite the online availability of data, analysis of this information in academic research is arduous. This article explores the application of supervised machine learning (SML) to overcome challenges associated with online data analysis. In SML classifiers are used to categorize and code binary data. Based on a case study of Dutch employees’ work-related tweets, this paper compares the coding performance of three classifiers, Linear Support Vector Machine, Naïve Bayes, and logistic regression. The performance of these classifiers is assessed by examining accuracy, precision, recall, the area under the precision-recall curve, and Krippendorf’s Alpha. These indices are obtained by comparing the coding decisions of the classifier to manual coding decisions. The findings indicate that the Linear Support Vector Machine and Naïve Bayes classifiers outperform the logistic regression classifier. This study also compared the performance of these classifiers based on stratified random samples and random samples of training data. The findings indicate that in smaller training sets stratified random training samples perform better than random training samples, in large training sets (n = 4000) random samples yield better results. Finally, the Linear Support Vector Machine classifier was trained with 4000 tweets and subsequently used to categorize 578,581 tweets obtained from 430 employees.

The probability of encountering an accidental DSM similarity in the presence of counter examples

The "combinatorial explosion" of DSM hypotheses can be explained by the occurrence of DSM similarities only due to an accidental coincidence of several non-essential properties of training examples. The standard mechanisms for reducing the number of hypotheses are increasing the border by the number of parent examples and prohibition of counter examples. Considering the Bernoulli trials (for each training example and each attribute) with success probability pj for the jth non-essential attribute, we model a random training sample set for the occurrence of accidental DSM similarity. The same method is used to specify a set of counter examples. For this model, we compute the generating function of an accidental resemblance to the b parent examples at m potential counter examples.

The use of single-date MODIS imagery for estimating large-scale urban impervious surface fraction with spectral mixture analysis and machine learning techniques

Urban impervious surface information is essential for urban and environmental applications at the regional/national scales. As a popular image processing technique, spectral mixture analysis (SMA) has rarely been applied to coarse-resolution imagery due to the difficulty of deriving endmember spectra using traditional endmember selection methods, particularly within heterogeneous urban environments. To address this problem, we derived endmember signatures through a least squares solution (LSS) technique with known abundances of sample pixels, and integrated these endmember signatures into SMA for mapping large-scale impervious surface fraction. In addition, with the same sample set, we carried out objective comparative analyses among SMA (i.e. fully constrained and unconstrained SMA) and machine learning (i.e. Cubist regression tree and Random Forests) techniques. Analysis of results suggests three major conclusions. First, with the extrapolated endmember spectra from stratified random training samples, the SMA approaches performed relatively well, as indicated by small MAE values. Second, Random Forests yields more reliable results than Cubist regression tree, and its accuracy is improved with increased sample sizes. Finally, comparative analyses suggest a tentative guide for selecting an optimal approach for large-scale fractional imperviousness estimation: unconstrained SMA might be a favorable option with a small number of samples, while Random Forests might be preferred if a large number of samples are available.

Abstract 239: Out-of-Hospital Cardiac Arrest Survival Predicted at 24 Hours: The Application and Validation of a Novel Clinical Score

Background : Despite 50 years of research, survival prognostication is still dependant on clinical status at 72 hours. We developed and tested the accuracy of a 24 hour, novel clinical score to predict the probability of survival post non-traumatic cardiac arrest. Methods : In 170 cardiac arrest survivors admitted between 2004-2010, a predictive model was developed. Variables used in univariate analyses to generate the prediction equation were: Initial Rhythm VF/VT, Return of Spontaneous Circulation (ROSC) ≤20 minutes, Brian Reflex Score (BRS)≥ 3 at 24 hours, Bystander CPR, and use of ≥3 doses of epinephrine. Multivariate logistic regression then retained significant predictors in a 50% random training sample, 84 patients (pts) and tested these in the remaining 50% validation sample, 86 pts. The diagnostic efficiency of the model was assessed using ROC analysis. Results : Of 170 pts studied, mean age was 63 years(SD±15.7), 37 pts (21.7%) survived to discharge. The initial rhythm was VF/VT in 38 (22%), 58 (34%) had ROSC≤ 20 minutes, 34 (20%) had a BRS ≥3 at 24 hours, and 80(47%) received Hypothermia. Mean length of stay was 5.8 days (SD±7.8).The final model retained three predictors: VF/VT, ROSC≤ 20 minutes and BRS≥ 3 at 24 hours. Table below presents the probability of survival with presence/ absence of all combinations of these 3 predictors. No single variable was robust in identifying the probability of survival (9%-17%). The presence of any 2 variables increases the probability to (55%-71%),however the presence of all 3 variables, in any patient increases the probability to 95%. The area under the ROC curve was 0.87 in the training sample and 0.84 in the validation sample. Conclusion : This study demonstrates that a score comprising 3 easily clinically, accessible variables correctly identifies survival following cardiac arrest. This score,like many other clinical scores, is incremental in predictive accuracy and warrant revalidation in larger cohort.

Comparing endmember selection techniques for accurate mapping of plant species and land cover using imaging spectrometer data

Building representative spectral libraries and quantitatively selecting a subset of spectra for mapping plant species and land cover/land use within remotely sensed imagery remain challenging for accurate classification. Multiple Endmember Spectral Mixture Analysis (MESMA) can be used for both classification and modeling fractional composition, and has been applied to map multiple biogeophysical variables. Our major objectives in this research were to 1) test a sampling design for building independent and representative training and validation spectral libraries; 2) compare endmember selection by a combination of two established techniques (count-based selection (CoB) and endmember average root mean square error (EAR)) with a recently introduced one (iterative endmember selection (IES)); and 3) develop and test a hybrid method, which combines the strengths of the previous two methods. We applied CoB/EAR, IES, and the new hybrid technique to mapping plant species and cover types in the Santa Ynez Mountains and Santa Barbara urban area, California, USA, using Airborne Visible Infrared Imaging Spectrometer (AVIRIS) data. For all endmember selection techniques, the number of selected endmembers varied across 25 random training samples. IES was consistently more accurate than CoB/EAR, but resulted in spectral libraries more than twice as large and failed to model rare species. The hybrid endmember selection technique resulted in the highest overall accuracy and kappa values and proved to be least sensitive to the random sampling protocols, but also produced the largest spectral libraries. A modified hybrid method, in which the number of endmembers selected was limited, produced the second highest accuracies, combining the strengths of the more parsimonious endmember selection by CoB/EAR with improved endmember selection by IES. Both IES and the hybrid methods selected endmembers that successfully classified a wide range of plant species and cover types, indicating their usefulness for these applications.

Exact Misclassification Probabilities for Plug-In Normal Quadratic Discriminant Functions: II. The Heterogeneous Case

We consider the problem of discriminating between two independent multivariate normal populations, N p ( μ 1, Σ 1) and N p ( μ 2, Σ 2), having distinct mean vectors μ 1 and μ 2 and distinct covariance matrices Σ 1 and Σ 2. The parameters μ 1, μ 2, Σ 1, and Σ 2 are unknown and are estimated by means of independent random training samples from each population. We derive a stochastic representation for the exact distribution of the “plug-in” quadratic discriminant function for classifying a new observation between the two populations. The stochastic representation involves only the classical standard normal, chi-square, and F distributions and is easily implemented for simulation purposes. Using Monte Carlo simulation of the stochastic representation we provide applications to the estimation of misclassification probabilities for the well-known iris data studied by Fisher ( Ann. Eugen. 7 (1936), 179–188); a data set on corporate financial ratios provided by Johnson and Wichern ( Applied Multivariate Statistical Analysis, 4th ed., Prentice–Hall, Englewood Cliffs, NJ, 1998); and a data set analyzed by Reaven and Miller ( Diabetologia 16 (1979), 17–24) in a classification of diabetic status.

Prediction off the left ventricular mass from the electrocardiogram in systemic hypertension

Although echocardiography provides a reliable method to determine left ventricular (LV) mass, it may not be available in all settings. Numerous electrocardiographic (ECG) criteria for the detection of LV hypertrophy have been developed, but few attempts have been made to predict the LV mass itself from the ECG. In a community-based survey program in the general population, 277 subjects were identified with untreated diastolic hypertension (diastolic blood pressure 95 to 115 mm Hg, 3 occasions) or isolated systolic hypertension (diastolic blood pressure < 95 mm Hg and systolic blood pressure 160 to 220 mm Hg, 3 occasions). All subjects underwent ECG and echocardiography on the same day. A multiple linear regression analysis was performed using a random training sample of the data set (n = 185). The independent variables included both ECG and non-ECG variables. The resulting model was used to predict the LV mass in the remainder of the data set, the validation sample (n = 92). Using sex, age, body surface area, the S-voltage in V 1 and V 4, and the duration of the terminal P in V 1 as independent variables, the model explained 45% of the variance (r = 0.67) in the training sample and 42% (r = 0.65) in the validation sample. This result exceeded that of 2 existing ECG models for LV mass (r = 0.40 and 0.41). The correlations between LV mass and combinations of ECG variables used for the detection of LV hypertrophy, such as the Sokolow-Lyon Voltage (r = 0.03) and the Cornell Voltage (r = 0.31), were comparatively low. In settings where echocardiography is not available or is too expensive and time-consuming, prediction of the LV mass from the ECG may offer a valuable alternative.

Rotation-invariant pattern recognition using optimum feature extraction

In this paper we consider the problem of detecting a target regardless of its orientation when it is known that the target must be from one of two classes. We assume significant random intraclass variability, a complication which requires techniques from statistical pattern recognition for amelioration. The Foley-Sammon transformation for selecting optimum features from random training samples is used to solve the problem.