Specific Training Data Research Articles

USPoint: Self-Supervised Interest Point Detection and Description for Ultrasound-Probe Motion Estimation During Fine-Adjustment Standard Fetal Plane Finding.

Ultrasound (US)-probe motion estimation is a fundamental problem in automated standard plane locating during obstetric US diagnosis. Most recent existing recent works employ deep neural network (DNN) to regress the probe motion. However, these deep regressionbased methods leverage the DNN to overfit on the specific training data, which is naturally lack of generalization ability for the clinical application. In this paper, we are back to generalized US feature learning rather than deep parameter regression. We propose a self-supervised learned local detector and descriptor, named USPoint, for US-probe motion estimation during the fine-adjustment phase of fetal plane acquisition. Specifically, a hybrid neural architecture is designed to simultaneously extract a local feature, and further estimate the probe motion. By embedding a differentiable USPoint-based motion estimation inside the proposed network architecture, the USPoint learns the keypoint detector, scores and descriptors from motion error alone, which doesn't require expensive human-annotation of local features. The two tasks, local feature learning and motion estimation, are jointly learned in a unified framework to enable collaborative learning with the aim of mutual benefit. To the best of our knowledge, it is the first learned local detector and descriptor tailored for the US image. Experimental evaluation on real clinical data demonstrates the resultant performance improvement on feature matching and motion estimation for potential clinical value. A video demo can be found online: https://youtu.be/JGzHuTQVlBs.

Knowledge graph augmented advanced learning models for commonsense reasoning

Machine learning is the key solution to many AI issues, but learning models rely heavily on specific training data. While a Bayesian setup can be used to incorporate some learning patterns with previous knowledge, those patterns can not access any organized world knowledge on requirements. The primary objective is to enable human-capable machines in ordinary everyday circumstances to estimate and make presumptions. In this paper we propose to respond to such common sense issues through a textual inference system with external, organized common sense graphs for explanatory inferences. The framework is based on a schematic map as a pair of questions and answers, a linked subgraph from the semantine to the symbolic space of knowledge-based external information. It displays a schematic map with a new network graphic module for information knowledge and performance with graph representations. LSTMs and graphical networks with a hierarchical attention-based direction are the basis of our model. It is flexible and understandable from the intermediate attention scores, leading to confident results. We also achieved state-of-the-art reliability on CommonsenseQA, a broad database of common sense reasoning utilizing ConceptNet as the only external tool for BERT-based models.

Injury and training history are associated with glenohumeral internal rotation deficit in youth tennis athletes

A glenohumeral internal rotation deficit (GIRD) of the shoulder, is associated with an increased risk of shoulder injuries in tennis athletes. The aim of the present study was to reveal the impact of 1) age, sex, specific training data (i.e. training volume, years of tennis practice, years of competitive play) and 2) upper extremity injuries on GIRD in youth competitive tennis athletes.A cross-sectional retrospective study design was adopted. Youth tennis players (n = 27, 12.6 ± 1.80 yrs., 18 male) belonging to an elite tennis squad were included. After documenting the independent variables (anthropometric data, tennis specific data and history of injury), the players were tested for internal (IR) and external (ER) shoulder rotation range of motion (RoM, [°]). From these raw values, the GIRD parameters ER/IR ratio and side differences and TRoM side differences were calculated. Pearson’s correlation analyses were performed to find potential associations of the independent variables with the GIRD outcomes.A significant positive linear correlation between the years of tennis training and IR side asymmetry occurred (p < .05). A significant negative linear relation between the years of tennis training and the ratio of ER to IR range of motion (RoM) in the dominant side (p < .05) was found. The analysis of covariance showed a significant influence of the history of injuries on IR RoM (p < .05).Injury and training history but not age or training volume may impact on glenohumeral internal rotation deficit in youth tennis athletes. We showed that GIRD in the dominant side in youth tennis players is progressive with increasing years of tennis practice and independent of years of practice associated with the history of injuries. Early detection of decreased glenohumeral RoM (specifically IR), as well as injury prevention training programs, may be useful to reduce GIRD and its negative consequences.

Light Actuation Based On Facial Mood Recognition

Current lighting technologies extend the options for changing the appearance of rooms and closed spaces, as such creating ambiences with an affective meaning. Using intelligence, these ambiences may instantly be adapted to the needs of the room’s occupant(s), possibly improving their well-being. In this paper, we set actuate lighting in our surrounding using mood detection. We analyze the mood of the person by Facial Emotion Recognition using deep learning model such as Convolutional Neural Network (CNN). On recognizing this emotion, we will actuate lighting in our surrounding in accordance with the mood. Based on implementation results, the system needs to be developed further by adding more specific data class and training data.

Real-world application of machine-learning-based fault detection trained with experimental data

Buildings are responsible for a large portion of the overall energy consumption. With the rising penetration of renewable energies, the heating and cooling demand of buildings will be increasingly satisfied by heat pumps. However, faults in the heat pump systems reduce energy efficiency or cause system failure, leading to an increased demand for primary energy. Hence, fault detection algorithms (FDA) are used to identify faults before system failure or efficiency deterioration occurs.With the rise of artificial intelligence and big data as well as more detailed monitoring systems, data-driven FDA have become a focus of research in recent years, showing promising results with acceptable effort. However, studies often use specific training data sets, thus generating FDAs adapted to a specific experimental system.In this paper, we investigate whether FDAs trained on a fault data set gathered with a laboratory heat pump system can be deployed on a real-world application system without the need for expensive modifications. We also investigate a big data approach, in which we use data collected over an extended period of time to train the FDAs.To this end, we use a data set kindly provided by the National Institute of Standards and Technology (NIST) containing data for typical heat pump failures measured on a specially outfitted air-water heat pump. From this data, we extract a series of features as input for the FDAs and evaluate the importance of those features for the FDAs. We train the algorithms to detect faults on the NIST data set, and transfer the fitted FDAs to our own measurement data.The results show that the trained FDAs perform very well on the NIST data set, but poorly on the real-world data set. We identify several reasons for the FDAs’ poor performance and derive mitigating actions. We believe that big data approaches for FDAs are facing several issues beyond the simple gathering of large data quantities, especially the labelling of occurred faults and completeness of the data set.

Rapid energy expenditure estimation for ankle assisted and inclined loaded walking

BackgroundEstimating energy expenditure with indirect calorimetry requires expensive equipment and several minutes of data collection for each condition of interest. While several methods estimate energy expenditure using correlation to data from wearable sensors, such as heart rate monitors or accelerometers, their accuracy has not been evaluated for activity conditions or subjects not included in the correlation process. The goal of our study was to develop data-driven models to estimate energy expenditure at intervals of approximately one second and demonstrate their ability to predict energetic cost for new conditions and subjects. Model inputs were muscle activity and vertical ground reaction forces, which are measurable by wearable electromyography electrodes and pressure sensing insoles.MethodsWe developed models that estimated energy expenditure while walking (1) with ankle exoskeleton assistance and (2) while carrying various loads and walking on inclines. Estimates were made each gait cycle or four second interval. We evaluated the performance of the models for three use cases. The first estimated energy expenditure (in Watts) during walking conditions for subjects with some subject specific training data available. The second estimated all conditions in the dataset for a new subject not included in the training data. The third estimated new conditions for a new subject.ResultsThe mean absolute percent errors in estimated energy expenditure during assisted walking conditions were 4.4%, 8.0%, and 8.1% for the three use cases, respectively. The average errors in energy expenditure estimation during inclined and loaded walking conditions were 6.1%, 9.7%, and 11.7% for the three use cases. For models not using subject-specific data, we evaluated the ability to order the magnitude of energy expenditure across conditions. The average percentage of correctly ordered conditions was 63% for assisted walking and 87% for incline and loaded walking.ConclusionsWe have determined the accuracy of estimating energy expenditure with data-driven models that rely on ground reaction forces and muscle activity for three use cases. For experimental use cases where the accuracy of a data-driven model is sufficient and similar training data is available, standard indirect calorimetry could be replaced. The models, code, and datasets are provided for reproduction and extension of our results.

Improving multilingual speech emotion recognition by combining acoustic features in a three-layer model

This study presents a scheme for multilingual speech emotion recognition. Determining the emotion of speech in general relies upon specific training data, and a different target speaker or language may present significant challenges. In this regard, we first explore 215 acoustic features from emotional speech. Second, we carry out speaker normalization and feature selection to develop a shared standard acoustic parameter set for multiple languages. Third, we use a three-layer model composed of acoustic features, semantic primitives, and emotion dimensions to map acoustics into emotion dimensions. Finally, we classify the continuous emotion dimensional values into basic categories by using the logistic model trees. The proposed approach was tested on Japanese, German, Chinese, and English emotional speech corpora. The recognition performance was examined and enhanced by cross-speaker and cross-corpus evaluation, and stressed the fact that our strategy is particularly suited for the task of multilingual emotion recognition even with a different speaker or language. The experimental results were found to be reasonably comparable with those of monolingual emotion recognizers as a reference.

Classification of motor imagery and execution signals with population-level feature sets: implications for probe design in fNIRS based BCI

Objective. The aim of this study was to introduce a novel methodology for classification of brain hemodynamic responses collected via functional near infrared spectroscopy (fNIRS) during rest, motor imagery (MI) and motor execution (ME) tasks which involves generating population-level training sets. Approach. A 48-channel fNIRS system was utilized to obtain hemodynamic signals from the frontal (FC), primary motor (PMC) and somatosensory cortex (SMC) of ten subjects during an experimental paradigm consisting of MI and ME of various right hand movements. Classification accuracies of random forest (RF), support vector machines (SVM), and artificial neural networks (ANN) were computed at the single subject level by training each classifier with subject specific features, and at the group level by training with features from all subjects for ME versus Rest, MI versus Rest and MI versus ME conditions. The performances were also computed for channel data restricted to FC, PMC and SMC regions separately to determine optimal probe location. Main results. RF, SVM and ANN had comparably high classification accuracies for ME versus Rest (%94, %96 and %98 respectively) and for MI versus Rest (%95, %95 and %98 respectively) when fed with group level feature sets. The accuracy performance of each algorithm in localized brain regions were comparable (>%93) to the accuracy performance obtained with whole brain channels (>%94) for both ME versus Rest and MI versus Rest conditions. Significance. By demonstrating the feasibility of generating a population level training set with a high classification performance for three different classification algorithms, the findings pave the path for removing the necessity to acquire subject specific training data and hold promise for a novel, real-time fNIRS based BCI system design which will be most effective for application to disease populations for whom obtaining data to train a classification algorithm is not possible.

Dataless Black-Box Model Comparison

In a time where the training of new machine learning models is extremely time-consuming and resource-intensive and the sale of these models or the access to them is more popular than ever, it is important to think about ways to ensure the protection of these models against theft. In this paper, we present a method for estimating the similarity or distance between two black-box models. Our approach does not depend on the knowledge about specific training data and therefore may be used to identify copies of or stolen machine learning models. It can also be applied to detect instances of license violations regarding the use of datasets. We validate our proposed method empirically on the CIFAR-10 and MNIST datasets using convolutional neural networks, generative adversarial networks and support vector machines. We show that it can clearly distinguish between models trained on different datasets. Theoretical foundations of our work are also given.

On better training the infinite restricted Boltzmann machines

The infinite restricted Boltzmann machine (iRBM) is an extension of the classic RBM. It enjoys a good property of automatically deciding the size of the hidden layer according to specific training data. With sufficient training, the iRBM can achieve a competitive performance with that of the classic RBM. However, the convergence of learning the iRBM is slow, due to the fact that the iRBM is sensitive to the ordering of its hidden units, the learned filters change slowly from the left-most hidden unit to right. To break this dependency between neighboring hidden units and speed up the convergence of training, a novel training strategy is proposed. The key idea of the proposed training strategy is randomly regrouping the hidden units before each gradient descent step. Potentially, a mixing of infinite many iRBMs with different permutations of the hidden units can be achieved by this learning method, which has a similar effect of preventing the model from over-fitting as the dropout. The original iRBM is also modified to be capable of carrying out discriminative training. To evaluate the impact of our method on convergence speed of learning and the model's generalization ability, several experiments have been performed on the binarized MNIST and CalTech101 Silhouettes datasets. Experimental results indicate that the proposed training strategy can greatly accelerate learning and enhance generalization ability of iRBMs.

Domain Adaptation in Steganalysis for the Spatial Domain

A scenario of domain adaptation (DA) in machine learning occurs when training and test data are drawn from some population with different distributions. In steganalysis, this scenario can arise when images used for training and testing come from different cameras, especially in blind detection. Although there has been some work in this area, it is still not clear that one can design a feasible detection scheme for all devices from one camera model. In this research, Spatial Rich Models (SRM) and ensemble classifiers have been applied for feature extraction and classification, respectively. After carefully collecting images from several camera models from mobile phones, with at least two devices for each model, we identify two measurable factors that affect detection: ISO speed and exposure time. This allows us to adapt the classifier from one device to a different one of the same model, even when images from the two devices are significantly different in visual appearance, by choosing specific training data. Our experiments show that a well-trained stego detector based on data from one source shows more adaptability to new target data if the training images have similar distributions of ISO speed and exposure time as the target images.

Detecting Multiple Coexisting Emotions in Microblogs with Convolutional Neural Networks

Analyzing human sentiments and emotions is a critical problem in cognitive computing. One fundamental task of sentiment analysis is to infer the sentiment polarity or emotion category of subjective text, such as microblogs. Most existing methods treat sentiment classification as a type of single-label supervised learning problem that classifies a microblog according to sentiment polarity or a single-labeled emotion. However, multiple fine-grained emotions may coexist in a single tweet or sentence of a microblog. We regard emotion detection in microblogs as a multi-label classification problem. First, we develop a graph-based algorithm to automatically build emotion lexicons, which are further utilized to construct distant-supervised corpora from massive microblog datasets. Then, a ranking-based multi-label convolutional neural network model (RM-CNN) that considers the order and relevance of labels is proposed to address emotion detection in microblogs. The RM-CNN model is pre-trained using the distant-supervised corpus and then fine-tuned using specific training data without the need for any manually designed features. Extensive experiments on two real-world datasets demonstrate substantial improvements of our proposed RM-CNN model over the state-of-the-art baseline methods in terms of multi-label classification metrics. We propose an effective RM-CNN model with a distant-supervised learning framework for detecting multiple coexisting emotions in the short text of microblogs.

Object-based urban structure type pattern recognition from Landsat TM with a Support Vector Machine

ABSTRACTThis study evaluates the potential of object-based image analysis in combination with supervised machine learning to identify urban structure type patterns from Landsat Thematic Mapper (TM) images. The main aim is to assess the influence of several critical choices commonly made during the training stage of a learning machine on the classification performance and to give recommendations for classifier-dependent intelligent training. Particular emphasis is given to assess the influence of size and class distribution of the training data, the approach of training data sampling (user-guided or random) and the type of training samples (squares or segments) on the classification performance of a Support Vector Machine (SVM). Different feature selection algorithms are compared and segmentation and classifier parameters are dynamically tuned for the specific image scene, classification task, and training data. The performance of the classifier is measured against a set of reference data sets from manual image interpretation and furthermore compared on the basis of landscape metrics to a very high resolution reference classification derived from light detection and ranging (lidar) measurements. The study highlights the importance of a careful design of the training stage and dynamically tuned classifier parameters, especially when dealing with noisy data and small training data sets. For the given experimental set-up, the study concludes that given optimized feature space and classifier parameters, training an SVM with segment-shaped samples that were sampled in a guided manner and are balanced between the classes provided the best classification results. If square-shaped samples are used, a random sampling provided better results than a guided selection. Equally balanced sample distributions outperformed unbalanced training sets.

Lazy Quantum clustering induced radial basis function networks (LQC-RBFN) with effective centers selection and radii determination

The Radial Basis Function Networks (RBFN) model has been successfully applied to different application scenarios as a universal approximator because of its simple architecture and online training capability. The approximation capability of RRBFN is greatly dependent on determination of the centers and the radii of the radial basis functions (RBFs) in the networks structure. Statistics-based centers determination approaches like K-means fail to capture and preserve the training data structure. In this paper, a new unsupervised RBFN construction methodology called Lazy Quantum Clustering induced Radial Basis Function Networks (LQC-RBFN) is proposed. It inherits the advantage of data structure learning and shows high robustness towards data distribution of Quantum Clustering (QC). At the same time, the controlling parameter can be determined arbitrarily without the requirement of precise calibration, and the minima search is done only once for a specific training data set. The centers and radii are selected based on the potential function generated by quantum assimilation, and the networks structure is adaptively updated incorporating the centers information. A series of application studies are presented to verify the effectiveness of the proposed LQC-RBFN model.

Accurate Prediction of Docked Protein Structure Similarity

One of the major challenges for protein-protein docking methods is to accurately discriminate nativelike structures. The protein docking community agrees on the existence of a relationship between various favorable intermolecular interactions (e.g. Van der Waals, electrostatic, desolvation forces, etc.) and the similarity of a conformation to its native structure. Different docking algorithms often formulate this relationship as a weighted sum of selected terms and calibrate their weights against specific training data to evaluate and rank candidate structures. However, the exact form of this relationship is unknown and the accuracy of such methods is impaired by the pervasiveness of false positives. Unlike the conventional scoring functions, we propose a novel machine learning approach that not only ranks the candidate structures relative to each other but also indicates how similar each candidate is to the native conformation. We trained the AccuRMSD neural network with an extensive dataset using the back-propagation learning algorithm. Our method achieved predicting RMSDs of unbound docked complexes with 0.4Å error margin.

Classifying the auditory P300 using mobile EEG recordings without calibration phase.

One of the major drawbacks in mobile EEG Brain Computer Interfaces (BCI) is the need for subject specific training data to train a classifier. By removing the need for supervised classification and calibration phase, new users could start immediate interaction with a BCI. We propose a solution to exploit the structural difference by means of canonical polyadic decomposition (CPD) for three-class auditory oddball data without the need for subject-specific information. We achieve this by adding average event-related-potential (ERP) templates to the CPD model. This constitutes a novel similarity measure between single-trial pairs and known-templates, which results in a fast and interpretable classifier. These results have similar accuracy to those of the supervised and cross-validated stepwise LDA approach but without the need for having subject-dependent data. Therefore the described CPD method has a significant practical advantage over the traditional and widely used approach.

Deep neural network acoustic models for spoken assessment applications

In this paper, we investigate the effectiveness of applying deep neural network hidden Markov models, or DNN-HMMs, for acoustic modeling in the context of educational applications. Specifically, we focus on spoken responses from non-native and child speech that tend to show great acoustic variability. We perform comprehensive experiments to compare the performance between traditional Gaussian mixture model (GMM)-HMMs and DNN-HMMs in three large language assessment datasets that contain various spoken tasks, classified broadly as constrained and open-ended tasks. Our experimental results suggest useful conclusions that can help guide the design of real-life educational applications. DNN-HMMs outperform conventional GMM-HMMs by a large margin for all spoken tasks commonly used in spoken assessment applications. In our experiments, DNN-HMMs trained using 25h of data can outperform GMM-HMMs trained with 6.7–9 times data. Specifically regarding overall performance, when all available training data were used (175, 227, 169h respectively), we achieved a relative word error rate decrease of 20.4% for adult English and 29.3% for child English, and a relative character error rate decrease of 14.3% for adult Chinese, when switching from GMMs to DNNs. In comparing between types of tasks, we notice that the more challenging open-ended tasks benefit significantly more than constrained item types by the use of DNN-HMMs. For open-ended tasks, having large amounts of training data is the key, as DNN-HMMs can take full advantage of the added training data and further push performance. In contrast, the performance of constrained spoken tasks saturates at around 25h of training data. At the same time, constrained spoken tasks require only a few hours of data (1 or 5h) to build well-performing acoustic models. This is an encouraging observation, that indicates the potential to build reliable spoken assessment applications based on constrained tasks, when few domain specific training data are available.

Recognition of Multi-Stroke Based Online Handwritten Gurmukhi Aksharas

In this paper, we have proposed Support Vector Machine (SVM) based efficient algorithm for Gurmukhi akshara formation/recognition. To train SVM classifier, a specific training data set has been used that forms an initial base to classify it. Experiment results have been obtained by using the software LibSVM where after preprocessing, coordinates were scaled from1 to 9 using the tool of LibSVM. In this experiment, 46,772 words were initially collected. In these words, 2,47,697 strokes were identified and further annotated. We tested the proposed methodology on 4,310 samples of Gurmukhi aksharas collected from different users. Testing results clearly reveal that for different combinations of Gurmukhi strokes and vowel/nasal, a high degree of accuracy has been achieved.

A simple method for assigning genomic grade to individual breast tumours

BackgroundThe prognostic value of grading in breast cancer can be increased with microarray technology, but proposed strategies are disadvantaged by the use of specific training data or parallel microscopic grading. Here, we investigate the performance of a method that uses no information outside the breast profile of interest.ResultsIn 251 profiled tumours we optimised a method that achieves grading by comparing rank means for genes predictive of high and low grade biology; a simpler method that allows for truly independent estimation of accuracy. Validation was carried out in 594 patients derived from several independent data sets. We found that accuracy was good: for low grade (G1) tumors 83- 94%, for high grade (G3) tumors 74- 100%. In keeping with aim of improved grading, two groups of intermediate grade (G2) cancers with significantly different outcome could be discriminated.ConclusionThis validates the concept of microarray-based grading in breast cancer, and provides a more practical method to achieve it. A simple R script for grading is available in an additional file. Clinical implementation could achieve better estimation of recurrence risk for 40 to 50% of breast cancer patients.

Clinical epidemiology and individualized medicine

Clinical epidemiology and individualized medicine