Neural Net Classifier Research Articles

Finger Tapping as a Behavioral Indicator of Presence in VR

Abstract The sense of spatial presence in virtual environments (VEs) is typically measured using self-report methods after task completion. Recent years have seen efforts to measure presence via physiological markers, with mixed results. Behavioral indicators might better track variations in presence over time. We propose finger tapping as an indicator of changing attentional demands corresponding to low and high presence conditions in VEs. Participants completed “virtual plank” (raised vs lowered to reflect high or low presence) and VR video (90- vs 360-degree) conditions while attempting to tap their fingers at a specified rate. Participants wore a data glove to measure finger taps, and we analyzed variations in the inter-tap intervals. We trained a neural net classifier to distinguish between low and high presence conditions using features derived from the positional finger-tap timeseries (e.g., power spectral density, autocorrelation). The trained network classified test epochs with a success rate of 80% for both video and plank conditions.

Benthic habitat mapping for estimating seagrass carbon stock across Takabonerate Islands, Indonesia

Data on benthic habitat cover are essential for monitoring and managing coastal areas, providing valuable ecosystem services for local communities. Moreover, seagrass extent is crucial for estimating carbon stock using Tier-1 calculations. On the other side, remote sensing technology has proven effective in mapping benthic habitat cover, including seagrass, over large local areas. Therefore, this research aims to map specific benthic habitats in a local area to obtain high-accuracy estimates of seagrass extent for carbon stock estimation. The study was conducted in the Takabonerate Islands, Indonesia, using Sentinel-2A imagery. The study involved field data processing, which enabled the mapping of eight benthic habitat classes using Sentinel-2A images, along with the steps for Sentinel-2A image processing for benthic habitat mapping. The benthic habitat map was produced following image correction steps, including atmospheric, sunglint, and water column corrections, followed by image classification using a Neural Net Classifier (NNC). Seagrass carbon stock estimation was performed using the Tier-1 equation from IPCC 2013. The study revealed a total benthic habitat area of 57,538.04 ha, with a seagrass area of 3127.42 ha for carbon stock estimation. The potential seagrass carbon stock in the study area was estimated to range from 337.76 GgC/ha to 31.27 GgC/ha using the Tier-1 equation. The overall accuracy of the benthic habitat map was 80.5 %, calculated using a confusion matrix table. The spatial information from this study can be used by local coastal managers and governments for baseline and monitoring purposes. The key interpretations from this study provide valuable findings that can be adapted and improved for future local mapping purposes.

Time series and mel frequency analyses of wet and dry cough signals: A neural net classification

Time series and mel frequency analyses of wet and dry cough signals: A neural net classification

Noise sensitivity and stability of deep neural networks for binary classification

A first step is taken towards understanding often observed non-robustness phenomena of deep neural net (DNN) classifiers. This is done from the perspective of Boolean functions by asking if certain sequences of Boolean functions represented by common DNN models are noise sensitive or noise stable, concepts defined in the Boolean function literature. Due to the natural randomness in DNN models, these concepts are extended to annealed and quenched versions. Here we sort out the relation between these definitions and investigate the properties of two standard DNN architectures, the fully connected and convolutional models, when initiated with Gaussian weights.

Hierarchical distinctions in the production and perception of nuclear tunes in American English

In Autosegmental-Metrical models of intonational phonology, different types of pitch accents, phrase accents, and boundary tones concatenate to create a set of phonologically distinct phrase-final nuclear tunes. This study asks if an eight-way distinction in nuclear tune shape in American English, predicted from the combination of two (monotonal) pitch accents, two phrase accents, and two boundary tones, is evident in speech production and in speech perception. F0 trajectories from a large-scale imitative speech production experiment were analyzed using bottom-up(k-means) clustering, neural net classification, GAMM modeling, and modeling of turning point alignment. Listeners’ perception of the same tunes is tested in a perceptual discrimination task and related to the imitation results. Emergent grouping of tunes in the clustering analysis, and related classification accuracy from the neural net, show a merging of some of the predicted distinctions among tunes whereby tune shapes that vary primarily in the scaling of final f0 are not reliably distinguished. Within five emergent clusters, subtler distinctions among tunes are evident in GAMMs and f0 turning point modeling. Clustering of individual participants’ production data shows a range of partitions of the data, with nearly all participants making a primary distinction between a class of High-Rising and Non-High-Rising tunes, and with up to four secondary distinctions among the non-Rising class. Perception results show a similar pattern, with poor pairwise discrimination for tunes that differ primarily, but by a small degree, in final f0, and highly accurate discrimination when just one member of a pair is in the High-Rising tune class. Together, the results suggest a hierarchy of distinctiveness among nuclear tunes, with a robust distinction based on holistic tune shape and poorly differentiated distinctions between tunes with the same holistic shape but small differences in final f0. The observed distinctions from clustering, classification, and perception analyses align with the tonal specification of a binary pitch accent contrast {H*, L*} and a maximally ternary {H%, M%, L%} boundary tone contrast; the findings do not support distinct tonal specifications for the phrase accent and boundary tone from the AM model. 

A Comparison between Supervised Classification Methods: Study Case on Land Cover Change Detection Caused by a Hydroelectric Complex Installation in the Brazilian Amazon

The Volta Grande do Xingu (VGX) in the Amazon Forest of Brazil was chosen to analyze the land use and land cover changes (LULCC) from 2000 to 2017, with the aim of assessing the most suitable classification method for the area. Three parametric (Mahalanobis distance, maximum likelihood and minimum distance) and three non-parametric (neural net, random forest and support vector machine) classification algorithms were tested in two Landsat scenes. The accuracy assessment was evaluated through a confusion matrix. Change detection of the landscape was analyzed through the post-classification comparison method. While maximum likelihood was more capable of highlighting errors in individual classes, support vector machine was slightly superior when compared with the other non-parametric options, these being the most suitable classifiers within the scope of this study. The main changes detected in the landscape were from forest to agro-pasture, from forest/agro-pasture to river, and from river to non-river, resulting in rock exposure. The methodology outlined in this research highlights the usefulness of remote sensing tools in follow-up observations of LULCC in the study area (with the possibility of application to the entire Amazon rainforest). Thus, it is possible to carry out adaptive management that aims to minimize unforeseen or underestimated impacts in previous stages of environmental licensing.

COVID-19 Incidence Prediction Model Based on Community Behavior With Neural Networks

Abstract
 BACKGROUNDS : The COVID-19 pandemic has created a global health emergency that requires a public health response to prevent the spread of the virus.
 AIM: The purpose of this study was to determine the prediction model for the incidence of COVID-19 based on community behavior.
 METHODS: This study used a cross-sectional study design. The study population was all people aged >18 years in Medan City and obtained a sample of 395 people with stratified random sampling technique. The research instrument used a questionnaire in google form, then, using Microsoft Office Excel, we transferred the data from the survey to a computer program. Furthermore, the data was analyzed using the neural networks method. Then the features importance will be calculated using the Random Forest with Mean Decrease Impurity (RF-MDI) method.
 RESULT: The results showed that based on the confusion matrix, the prediction value for those who did not suffer from COVID-19 was correct from negative data = 8, the correct prediction value for COVID-19 from positive data = 8. While the incorrect prediction value for machines that predicted negative results but the actual data was positive = 2, and predicts a positive result but the actual data is negative = 4. Thus, based on the neural net classification method, the accuracy value is 72%. The results of this study indicate that poor preventive behavior by the community greatly affects the spread of COVID-19 cases.
 CONCLUSION: Poor community behavior, such as not limiting their interaction/contact with other people, not exercising frequently, leaving the house without keeping a safe distance, and not washing hands regularly, can all impact COVID-19 transmission in the community
 Keywords: Behavior Prediction Model, COVID-19 Incidence, Neural Network

Robustness verification of ReLU networks via quadratic programming

Neural networks are known to be sensitive to adversarial perturbations. To investigate this undesired behavior we consider the problem of computing the distance to the decision boundary (DtDB) from a given sample for a deep neural net classifier. In this work we present a procedure where we solve a convex quadratic programming (QP) task to obtain a lower bound on the DtDB. This bound is used as a robustness certificate of the classifier around a given sample. We show that our approach provides better or competitive results in comparison with a wide range of existing techniques.

507 Rational Radiomic Design for Stepwise Diagnosis of Posterior Fossa Pediatric Tumors

INTRODUCTION: Medulloblastoma (MB), pilocytic astrocytoma (PA), and ependymoma (EP) compose the majority of posterior fossa (PF) pediatric tumors and share similar features on MRI. Pre-operative anticipation of pathology can inform the surgical approach, extent of resection, and potential complications. METHODS: We extracted 1800 Image features from T2- and gadolinium-enhanced T1-weighted images in a multinational cohort of 274 MB, 156 PA, and 97 EP. We designed a two-step classifier—first ruling out PA with a three-way classifier, and next distinguishing MB from EP with a binary classifier. For each step, we selected the best performing classifier model from six candidates following LASSO-feature reduction. Final multi-class classifier performance was measured on a holdout test set with the micro-averaged F1-score. RESULTS: Optimal diagnostic performance was achieved using two decision steps, each with their respective feature set and classifier method. An initial three-way (MB, PA, and EP) logistic regression classifier exhibited a micro-averaged F1-score of 0.739. From a reduced feature set of 61 features, T2-Uniformity and T1-Contrast were the most relevant for distinguishing PA. A subsequent two-way neural net classifier distinguished MB from EP with F1-score 0.9189. In the second reduced feature set of 39 features, T2-Sphericity and T1-Flatness were most relevant. Performing the two classifiers sequentially for MB, PA, and EP classification produced a micro-averaged F1-score of 0.9179. CONCLUSION: A sequential radiomic improved upon a single-step radiographic classifier for pediatric PF tumors. Strong overlap between MB and EP prevented a single-step classifier from distinguishing all three pathologies at once. PA-relevant features aligned with avid gadolinium-enhancement and non-enhancing cystic components. MB and EP-related features correlated to how each conform within and extrude from the fourth ventricle, respectively.

Active monitoring of powder bed fusion process by training an artificial neural net classifier on layer-by-layer surface laser profilometry data

Active monitoring of powder bed fusion process by training an artificial neural net classifier on layer-by-layer surface laser profilometry data

Machine Assist for Pediatric Posterior Fossa Tumor Diagnosis: A Multinational Study.

Clinicians and machine classifiers reliably diagnose pilocytic astrocytoma (PA) on magnetic resonance imaging (MRI) but less accurately distinguish medulloblastoma (MB) from ependymoma (EP). One strategy is to first rule out the most identifiable diagnosis. To hypothesize a sequential machine-learning classifier could improve diagnostic performance by mimicking a clinician's strategy of excluding PA before distinguishing MB from EP. We extracted 1800 total Image Biomarker Standardization Initiative (IBSI)-based features from T2- and gadolinium-enhanced T1-weighted images in a multinational cohort of 274MB, 156 PA, and 97 EP. We designed a 2-step sequential classifier - first ruling out PA, and next distinguishing MB from EP. For each step, we selected the best performing model from 6-candidate classifier using a reduced feature set, and measured performance on a holdout test set with the microaveraged F1 score. Optimal diagnostic performance was achieved using 2 decision steps, each with its own distinct imaging features and classifier method. A 3-way logistic regression classifier first distinguished PA from non-PA, with T2 uniformity and T1 contrast as the most relevant IBSI features (F1 score 0.8809). A 2-way neural net classifier next distinguished MB from EP, with T2 sphericity and T1 flatness as most relevant (F1 score 0.9189). The combined, sequential classifier was with F1 score 0.9179. An MRI-based sequential machine-learning classifiers offer high-performance prediction of pediatric posterior fossa tumors across a large, multinational cohort. Optimization of this model with demographic, clinical, imaging, and molecular predictors could provide significant advantages for family counseling and surgical planning.

Efficient Estimation of the ANOVA Mean Dimension, with an Application to Neural Net Classification

The mean dimension of a black box function of $d$ variables is a convenient way to summarize the extent to which it is dominated by high or low order interactions. It is expressed in terms of $2^d-...

PATH-22. COMPARISON OF SUPERVISED CLASSIFICATION METHODS FOR CENTRAL NERVOUS SYSTEM TUMORS BASED ON DNA-METHYLATION

Classification of brain tumors using methylation profiling is an important diagnostic advance, reducing subjectivity and improving interpretability of clinical outcome data. Despite the recognized value of methylation profiling in the clinical laboratory, the performance characteristics of different supervised classification models has not been directly compared. We developed 3 methods using methylation profiles to classify CNS tumors: an exact bootstrap k-nearest neighbor (kNN), a multi-layer perceptron neural net (NN), and a random forest classifier (RF). We trained these methods on the publicly available CNS tumor reference cohort (GSE90496) with 2,801 profiles and 91 classes. We evaluated the performance of these methods by leave-out-25% cross-validation. The relative performance of these methods were evaluated in terms of accuracy, precision, and recall for class or class family. The kNN, RF, and NN classifier had an estimate error rate of 10.74%, 4.01%, and 1.89%, respectively for class prediction and an error rate for family prediction of 5.97%, 0.90%, and 0.6%, respectively. At perfect recall for class assignment, the RF and kNN had a precision of 0.96 and 0.89 while the NN reached 0.98. For family assignment, the precision for the three classifiers was almost 1.0 with recall of nearly 0.8. At the recall rate of 1.0, the precision dropped to 0.94, 0.991 and 0.994 for kNN, RF, and NN, respectively. Overall, the NN showed improved performance metrics compared to the kNN and RF in CNS tumor classification for both class and class family assignment.

Leveraging Very-High Spatial Resolution Hyperspectral and Thermal UAV Imageries for Characterizing Diurnal Indicators of Grapevine Physiology

Efficient and accurate methods to monitor crop physiological responses help growers better understand crop physiology and improve crop productivity. In recent years, developments in unmanned aerial vehicles (UAV) and sensor technology have enabled image acquisition at very-high spectral, spatial, and temporal resolutions. However, potential applications and limitations of very-high-resolution (VHR) hyperspectral and thermal UAV imaging for characterization of plant diurnal physiology remain largely unknown, due to issues related to shadow and canopy heterogeneity. In this study, we propose a canopy zone-weighting (CZW) method to leverage the potential of VHR (≤9 cm) hyperspectral and thermal UAV imageries in estimating physiological indicators, such as stomatal conductance (Gs) and steady-state fluorescence (Fs). Diurnal flights and concurrent in-situ measurements were conducted during grapevine growing seasons in 2017 and 2018 in a vineyard in Missouri, USA. We used neural net classifier and the Canny edge detection method to extract pure vine canopy from the hyperspectral and thermal images, respectively. Then, the vine canopy was segmented into three canopy zones (sunlit, nadir, and shaded) using K-means clustering based on the canopy shadow fraction and canopy temperature. Common reflectance-based spectral indices, sun-induced chlorophyll fluorescence (SIF), and simplified canopy water stress index (siCWSI) were computed as image retrievals. Using the coefficient of determination (R2) established between the image retrievals from three canopy zones and the in-situ measurements as a weight factor, weighted image retrievals were calculated and their correlation with in-situ measurements was explored. The results showed that the most frequent and the highest correlations were found for Gs and Fs, with CZW-based Photochemical reflectance index (PRI), SIF, and siCWSI (PRICZW, SIFCZW, and siCWSICZW), respectively. When all flights combined for the given field campaign date, PRICZW, SIFCZW, and siCWSICZW significantly improved the relationship with Gs and Fs. The proposed approach takes full advantage of VHR hyperspectral and thermal UAV imageries, and suggests that the CZW method is simple yet effective in estimating Gs and Fs.

Early detection and follow-up of colorectal neoplasia based on faecal volatile organic compounds.

Early detection and removal of colorectal cancer (CRC) and advanced adenomas (AAs) decreases the incidence of and mortality from the disease. We aimed to evaluate the potential of faecal volatile organic compounds (VOCs) for detection and follow-up of colorectal adenoma using advanced electronic nose technology. This was a prospective multi-centre case-control cohort including two district hospitals and one tertiary referral hospital. Patients undergoing colonoscopy were instructed to collect a faecal sample prior to bowel cleansing and were included in the study when CRC, AAs, large adenomas (LAs; 0.5-1.0cm), small adenomas (SAs; 0.1-0.5cm) or no endoscopic abnormalities (controls) were observed. Patients undergoing polypectomy and controls were asked for a second sample after 3months. Faecal VOCs were measured with gas chromatography-ion mobility spectrometry. Random forest, support vector machine, Gaussian process and neural net classification were used to evaluate accuracy. In total, 14 patients with CRC, 64 with AAs, 69 with LAs, 127 with SAs and 227 controls were included. A second sample was collected from 32 polypectomy patients and 32 controls. Faecal VOCs discriminated CRC and adenomas from control [AUC (95% CI): CRC vs control 0.96 (0.89-1); AA vs control 0.96 (0.93-1); LA vs control 0.96 (0.92-0.99); SA vs control 0.96 (0.94-0.99)]. There were no significant differences between CRC and adenoma groups. Patients with adenomas and controls were discriminated prior to polypectomy, whereas 3months after polypectomy VOC profiles were similar [T0 adenoma vs control 0.98 (0.95-1); T1 adenoma vs control 0.55 (0.40-0.69)]. Faecal VOC profiles may be useful for early detection of CRC and adenomas and the timing of polyp surveillance as polypectomy led to a normalization of the VOC profile.

Enhancing the Performance of Semantic Search in Bengali using Neural Net and other Classification Techniques

To know the information from the internet searching is one of the most important part for any user. In case of ‘Syntactic Search’ keyword based matching technique is used. Search accuracy is improved applying the filter like location, preference, user-history etc. However, it can happen that the user query or question and the best available answer or result in the internet domain has no terms in common or ignorable number of terms is common. In such case syntactic search cannot give the desired output. The role of ‘Semantic Search’ becomes prevalent in this scenario. The execution of semantic search faces challenge due to unavailability of resources like WordNet, Ontology, Annotation etc. An end to end algorithm is described to improve the accuracy of the semantic search in this work. Four classification techniques are used. They are ANN, Decision Tree, SVM and Naïve Bayes. Dataset is provided from the TDIL project of the Ministry of Electronics and IT, Govt. of India. The repository contains 86 categories of text having more than a million sentences. After getting the impressive result for the Bengali language test run was done for other Indian languages and a very good result is achieved. This research is extremely useful for the automatic question answering system, semantic similarity analysis, e-governance and m- governance.

Early Detection of the Risk of Developing Psychiatric Disorders: A Study of 461 Chinese University Students under Chronic Stress

Chronic stress, a characteristic of modern time, has a significant impact on general health. In the context of psychiatric disorders, insufficient coping behavior under chronic stress has been linked to higher rates of (1) depressive symptoms among subjects of the general population, (2) relapse among patients under treatment for clinical depression, and (3) negative symptoms among subjects with an elevated vulnerability to psychosis. In this normative study we assessed basic coping behavior among 461 Chinese freshman university students along with their consumption behavior and general health in terms of regular exercises, physical health, psychosomatic disturbances, and mental health. The assessments relied on two instruments that have already demonstrated their capability of (1) reliably detecting insufficient coping behavior under chronic stress and (2) reliably quantifying the interrelation between coping behavior and mental health in the Western world. Thus, we aimed to complement existing data and to develop a generally available, socioculturally independent tool that can be used for the early detection of subjects with an elevated risk of mental health problems. Structural analyses yielded essentially the same scales “activity” and “defeatism” as previous studies on 2,500 students from Switzerland, Italy, Spain, the USA, and Argentina. These scales explained 74.3% of the observed variance in coping behavior among the 461 Chinese students. We found highly significant correlations (p < 0.0001) between the “defeatism” scale on the one hand, and the scales “regular use of medicine,” “psychosomatic disturbances,” and “impaired mental health” on the other. Particularly intriguing was the finding that a neural net classifier could be constructed to identify students with the highest contributions to the interrelation between “coping behavior” and “mental health,” yielding a correlation coefficient as high as r = 0.597 for the respective subgroup. Based on the normative data, an online tool for risk assessments was developed with immediate feedback to users. This study provided another piece of evidence regarding the close link between basic coping behavior and mental health, across cultures and ethnicities. In consequence, our approach to quantifying basic coping behavior, along with other risk factors, can be expected to clear the way for an “early” detection of students with an elevated risk of stress-related mental health problems, nota bene prior to the development of clinically relevant symptoms. The socioeconomic impact of the potential prevention of depressive ­disorders, and psychiatric disorders in general, may be enormous.

Confidence-based voting procedure for combining fuzzy systems and neural networks

In this study the confidence-based voting of neural net classifier and fuzzy logic based classifiers is proposed. In this method, for the cases when the fuzzy system is confident enough in its decision, i.e. when the membership value is large enough, fuzzy system makes the decision, otherwise, the neural net is applied. This allows classifying most of the objects by explainable interpretable fuzzy system, while using the more accurate neural network for the most difficult cases. The experiments are performed on a set of test datasets, and two problems of identifying the emotional state of a person using the data collected by non-contact vital Doppler sensors. The results show that this setup allows not only improving the classification quality, but also allows to explain the classification process the explanation of the classifier functioning.

Machine learning approach to pulse shape discrimination in liquid noble gas detectors

Study of the event classification in liquid noble gas detectors is presented. The discrimination between different events usually done by prompt light fraction. To tackle this task we adopt the neural net classifier and use pulse shape as an input feature. The main difficulty comes from low-energy events that are difficult to separate. This is important because these events provide a background for dark matter searches. We find that neural networks are suitable for this task.

The influence of land use and land cover change on landslide susceptibility: a case study in Zhushan Town, Xuan'en County (Hubei, China)

Abstract. Land use and land cover change can increase or decrease landslide susceptibility (LS) in the mountainous areas. In the hilly and mountainous part of southwestern China, land use and land cover change (LUCC) has taken place in the last decades due to infrastructure development and rapid economic activities. This development and activities can worsen the slope susceptible to sliding due to mostly the cutting of slopes. This study, taking Zhushan Town, Xuan'en County, as the study area, aims to evaluate the influence of land use and land cover change on landslide susceptibility at a regional scale. Spatial distribution of landslides was determined in terms of visual interpretation of aerial photographs and remote sensing images, supported by field surveys. Two types of land use and land cover (LUC) maps, with a time interval covering 21 years (1992–2013), were prepared: the first was obtained by the neural net classification of images acquired in 1992 and the second by the object-oriented classification of images in 2002 and 2013. Landslide-susceptible areas were analyzed using the logistic regression model (LRM) in which six influencing factors were chosen as the landslide susceptibility indices. In addition, the hydrologic analysis method was applied to optimize the partitioning of the terrain. The results indicated that the LUCC in the region was mainly the transformation from the grassland and arable land to the forest land, which is increased by 34.3 %. An increase of 1.9 % is shown in the area where human engineering activities concentrate. The comparison of landslide susceptibility maps among different periods revealed that human engineering activities were the most important factor in increasing LS in this region. Such results emphasize the requirement of a reasonable land use planning activity process.