Data Machine Learning Algorithms Research Articles

Focused active learning for histopathological image classification

Active Learning (AL) has the potential to solve a major problem of digital pathology: the efficient acquisition of labeled data for machine learning algorithms. However, existing AL methods often struggle in realistic settings with artifacts, ambiguities, and class imbalances, as commonly seen in the medical field. The lack of precise uncertainty estimations leads to the acquisition of images with a low informative value. To address these challenges, we propose Focused Active Learning (FocAL), which combines a Bayesian Neural Network with Out-of-Distribution detection to estimate different uncertainties for the acquisition function. Specifically, the weighted epistemic uncertainty accounts for the class imbalance, aleatoric uncertainty for ambiguous images, and an OoD score for artifacts. We perform extensive experiments to validate our method on MNIST and the real-world Panda dataset for the classification of prostate cancer. The results confirm that other AL methods are ‘distracted’ by ambiguities and artifacts which harm the performance. FocAL effectively focuses on the most informative images, avoiding ambiguities and artifacts during acquisition. For both experiments, FocAL outperforms existing AL approaches, reaching a Cohen’s kappa of 0.764 with only 0.69% of the labeled Panda data.

Exploring classical machine learning for identification of pathological lung auscultations

The use of machine learning in biomedical research has surged in recent years thanks to advances in devices and artificial intelligence. Our aim is to expand this body of knowledge by applying machine learning to pulmonary auscultation signals. Despite improvements in digital stethoscopes and attempts to find synergy between them and artificial intelligence, solutions for their use in clinical settings remain scarce. Physicians continue to infer initial diagnoses with less sophisticated means, resulting in low accuracy, leading to suboptimal patient care. To arrive at a correct preliminary diagnosis, the auscultation diagnostics need to be of high accuracy. Due to the large number of auscultations performed, data availability opens up opportunities for more effective sound analysis. In this study, digital 6-channel auscultations of 45 patients were used in various machine learning scenarios, with the aim of distinguishing between normal and abnormal pulmonary sounds. Audio features (such as fundamental frequencies F0-4, loudness, HNR, DFA, as well as descriptive statistics of log energy, RMS and MFCC) were extracted using the Python library Surfboard. Windowing, feature aggregation, and concatenation strategies were used to prepare data for machine learning algorithms in unsupervised (fair-cut forest, outlier forest) and supervised (random forest, regularized logistic regression) settings. The evaluation was carried out using 9-fold stratified cross-validation repeated 30 times. Decision fusion by averaging the outputs for a subject was also tested and found to be helpful. Supervised models showed a consistent advantage over unsupervised ones, with random forest achieving a mean AUC ROC of 0.691 (accuracy 71.11%, Kappa 0.416, F1-score 0.675) in side-based detection and a mean AUC ROC of 0.721 (accuracy 68.89%, Kappa 0.371, F1-score 0.650) in patient-based detection.

Technique for preparation of large data for machine learning algorithms to generate intrusion detection system

Machine Learning has become the norm for creating models that predict and detect security breaches in computer systems. Modelling of machine learning systems requires a huge amount of data for training and testing the machine learning algorithm. The present research is concerned with preparing the data prior to feeding it to machine learning algorithms. The necessity for such preparation occurs due to the physical limitations of spreadsheet applications (with regard to the number of records they can handle). This research describes a simple shell scripting approach to combine, extract, and weed out unwanted records and finally prepare the data for feeding to machine learning algorithms.

Touch events and human activities for continuous authentication via smartphone

The security of modern smartphones is related to the combination of Continuous Authentication approaches, Touch events, and Human Activities. The approaches of Continuous Authentication, Touch Events, and Human Activities are silent to the user but are a great source of data for Machine Learning Algorithms. This work aims to develop a method for continuous authentication while the user is sitting and scrolling documents on the smartphone. Touch Events and Smartphone Sensor Features (from the well-known H-MOG Dataset) were used with the addition, for each sensor, of the feature called Signal Vector Magnitude. Several Machine Learning Models have been considered with different experiment setups, 1-class, and 2-class, for evaluation. The results show that the 1-class SVM achieves an accuracy of 98.9% and an F1-score of 99.4%, considering the selected features and the feature Signal Vector Magnitude very significant.

Counting Activities Using Weakly Labeled Raw Acceleration Data: A Variable-Length Sequence Approach with Deep Learning to Maintain Event Duration Flexibility.

This paper presents a novel approach for counting hand-performed activities using deep learning and inertial measurement units (IMUs). The particular challenge in this task is finding the correct window size for capturing activities with different durations. Traditionally, fixed window sizes have been used, which occasionally result in incorrectly represented activities. To address this limitation, we propose segmenting the time series data into variable-length sequences using ragged tensors to store and process the data. Additionally, our approach utilizes weakly labeled data to simplify the annotation process and reduce the time to prepare annotated data for machine learning algorithms. Thus, the model receives only partial information about the performed activity. Therefore, we propose an LSTM-based architecture, which takes into account both the ragged tensors and the weak labels. To the best of our knowledge, no prior studies attempted counting utilizing variable-size IMU acceleration data with relatively low computational requirements using the number of completed repetitions of hand-performed activities as a label. Hence, we present the data segmentation method we employed and the model architecture that we implemented to show the effectiveness of our approach. Our results are evaluated using the Skoda public dataset for Human activity recognition (HAR) and demonstrate a repetition error of ±1 even in the most challenging cases. The findings of this study have applications and can be beneficial for various fields, including healthcare, sports and fitness, human-computer interaction, robotics, and the manufacturing industry.

On Possibilities of Learning the Meta Data for Machine Learning Algorithms

On Possibilities of Learning the Meta Data for Machine Learning Algorithms

A Novel Method for Fast Generation of 3D Objects from Multiple Depth Sensors

Abstract Scanning real 3D objects face many technical challenges. Stationary solutions allow for accurate scanning. However, they usually require special and expensive equipment. Competitive mobile solutions (handheld scanners, LiDARs on vehicles, etc.) do not allow for an accurate and fast mapping of the surface of the scanned object. The article proposes an end-to-end automated solution that enables the use of widely available mobile and stationary scanners. The related system generates a full 3D model of the object based on multiple depth sensors. For this purpose, the scanned object is marked with markers. Markers type and positions are automatically detected and mapped to a template mesh. The reference template is automatically selected for the scanned object, which is then transformed according to the data from the scanners with non-rigid transformation. The solution allows for the fast scanning of complex and varied size objects, constituting a set of training data for segmentation and classification systems of 3D scenes. The main advantage of the proposed solution is its efficiency, which enables real-time scanning and the ability to generate a mesh with a regular structure. It is critical for training data for machine learning algorithms. The source code is available at https://github.com/SATOffice/improved_scanner3D.

Synthetic data generation for machine learning model training for energy theft scenarios using cosimulation

AbstractTechnical and non‐technical losses in distribution circuits result in significant economic costs to power utilities. One type of non‐technical loss is energy theft by various means including illegal tapping of feeders, bypassing the meter, and billing fraud. These losses are usually hard to detect, and can remain undetected for long periods of time. Machine learning models have been proven effective in detecting these conditions, but rely on the availability of large, good‐quality training data sets. The problem is exacerbated by the imbalanced nature of data related to these conditions—energy theft, though costly, is very rare. The available data sets generally have very few samples of theft with most of the data pertaining to normal operation. Such data sets are generally not suitable to train machine learning models. In this paper, an overview of energy theft detection techniques, the challenges with their data needs, and the limitations of current techniques to bridge such data limitations is presented. A co‐simulation framework is proposed to generate reliable training data for machine learning algorithms for theft detection. An example scenario is presented and a machine learning model is built to detect certain kinds of energy theft.

A post processing pipeline to prepare raw data for machine learning algorithms in cardiac magnetic resonance imaging

Abstract Funding Acknowledgements Type of funding sources: None. Background Artificial Intelligence is an emergent tool in clinical practice for post processing of medical images. Machine Learning (ML) pipelines are created for data of interest extraction and algorithm application. A common issue in data extraction is represented by noisy datasets, like those of CMR studies, characterized by multiple images, acquired by different techniques, axis orientation and contrast timing. Purpose A ML pipeline for extraction of LGE images from raw DICOM data is presented. Additionally, steps for normalization of image number and automatically heart localization are outlined. Methods and Results 642 consecutive CMR studies were analyzed. Pipeline, Part 1. By looking at the metadata in raw files, ‘SequenceName’ tag was used to discard cine images, ‘ScanningSequence’ tag to select Gradient Recall and Inversion Recovery techniques (Inversion Time > 100 ms), ‘SequenceVariant’ tag to discard Steady State images (See Fig. 1). Orientation of the major axis was computed and ‘Axial’ or ‘Coronal’ images removed. Scans were grouped together by image orientation (requesting a min and max number of elements per group) and only the group with the largest number of files was selected. Finally, DICOMs were grouped by image shape (demanding a min number of elements), and only the series with the highest resolution was retained. Then, for each subject, the extracted series consists of a 3D-array (N,H,W), with N number of slices, and (H,W) image resolution. The attributes were not homogeneous between subjects. Pipeline, Part 2. Given a desired final number of slices and resolution, the 3D-array was reshaped through a spline interpolation. In order to have a focus on the heart, a Region of Interest (ROI) extractor was implemented, based on a YOLO network for object detection. The network was applied on all the slices (Fig. 2); then the images were cropped by keeping the largest bounding box. This step allowed us to remove the background by only selecting the relevant ROIs. To manage the data more easily, images were saved as a NumPy Array, while other useful Dicom metadata (e.g. weight, age, …) were stored using the json standard. At the end of the ML pipeline, images can be reduced to a common resolution and forwarded to ML algorithms. By using this pipeline, a great amount of information not needed for LGE analysis can be discarded, granting a significant reduction in terms of data storage. In our series, the original dataset extended for about 200 GB; by requesting 10 slices per subject with a resolution of 128 by 128 pixels (also extracting heart ROI) the final dimension was reduced to 108 MB. Conclusions In this work, we presented a post-processing pipeline for CRM images and LGE analysis. Given in input raw CRM the pipeline is able to (i) remove unuseful data, (ii) extract heart ROIs storing also Dicom metadata, (iii) normalize slices and image resolution, and (iv) store the processed CRM into ready-format for ML techniques. Fig.1:Pipeline Schematic RepresentationFig.2:YOLO Heart Extraction

Embedded reproduction in platform data work

ABSTRACT This paper focuses on the experiences of Latin American data workers who annotate data for machine learning algorithms through labor platforms. It introduces the notion of ‘embedded reproduction’: the relationship between embeddedness, the degree to which non-economic institutions and their social environment constrain socioeconomic activity, and social reproduction, or the activities that nurture, maintain, and regenerate the workforce. The analysis of 38 interviews with platform workers suggests they are situated in a highly disembedded market due to the lack of regulations on the data production process, giving free rein to platforms to set rules to their detriment. This article explores how this disembeddedness shapes social reproduction by studying three forms of collective social support received by workers: from family members, neighbors and local communities, and online groups. The support of these networks is primarily local, depends on high levels of trust, and is gendered. These findings suggest that platform data work is unsustainable from an embedded reproductive perspective since platform intermediation leads workers and local communities to carry out the social and economic risks associated with this form of gig work. This research invites a dialogue between the embeddedness framework with social reproduction as well as a consideration of the importance of nature and natural resources in the study of social environments.

A comparative analysis of artificial neural network and support vector machine for online transient stability prediction considering uncertainties

ABSTRACT Power system transient stability is an integral part of power system planning and operation. Conventional approaches to assess transient stability are time consuming and hence, are not suitable for online application. Thus, this paper presents a comparative analysis of two different machine learning (ML) algorithms, i.e. artificial neural network (ANN) and support vector machine (SVM), for online transient stability prediction, considering various uncertainties (load, network topology, fault type, fault location, and fault clearing time). Time-domain simulations were conducted using DIgSILENT PowerFactory software for obtaining the training data for ML algorithms. MATLAB was used to apply the ML algorithms (ANN and SVM), and to draw a comparison between them. The classification accuracy of ANN and SVM was found to be 98.4% and 94.4% (fixed topology), respectively. The classification accuracy of ANN and SVM was found to be 97.7% and 93.2% (topology change), respectively. The training time of ANN was less than that of SVM (for both topology cases). These results for the IEEE 14-bus system demonstrated that both ANN and SVM can rapidly estimate the transient stability, considering uncertainties, with a reasonable accuracy; however, ANN outperformed SVM as its classification performance and computational performance was determined to be superior.

Applying Convolutional Neural Networks to Neuroimaging Classification Tasks: A Practical Guide in Python.

In this chapter, we describe the process of obtaining medical imaging data and its storage protocol. The authors also explain in a step-by-step approach how to extract and prepare the medical imaging data for machine learning algorithms. And finally, the process of building and assessing a convolutional neural network for medical imaging data is illustrated.

A Realistic Radar Ray Tracing Simulator for Large MIMO-Arrays in Automotive Environments

With the rise of high resolution multiple input multiple output (MIMO) systems, radar became an important sensor in the development of Advanced Driver Assistance Systems (ADAS) and autonomous driving applications. Autonomous driving will rely strongly on artificial intelligence. Since most modern classification algorithms are based on neural networks, they require huge amounts of data to perform well, especially in unexpected traffic situations. Radar sensor simulation can potentially produce a great variety of training data for machine learning algorithms, which makes it an important cornerstone in the development of ADAS. Furthermore, with radar simulators, different antenna configurations and various edge cases can be simulated. In this work, a versatile ray tracing toolchain based on the shoot and bouncing rays (SBR) approach is presented. The program is able to simulate complex urban environments including realistic clutter, by utilizing simplistic reflection models. The program does not only produce realistic radar images, but also generates camera-like images using the same materials. Furthermore, this work deals with the adaption of the SBR method to radar sensors with an arbitrary number of transmit (TX)- and receive (RX) antennas, which enables the simulation of large MIMO arrays. A novel performance optimization approach is proposed for large numbers of TX antennas, which reduces the runtime dramatically. The quality of the simulation is verified by measuring a complex and realistic scenario with a high resolution automotive MIMO radar. Also, a study of the effect on quality and runtime is being investigated for various optimization approaches, including the proposed method.

DISEMBEDDEDNESS IN MACHINE LEARNING DATA WORK

Firms and research organizations require humans to annotate raw data to make it compatible with machine learning algorithms. These tasks are often outsourced to individuals worldwide through labor platforms or infrastructures that serve as marketplaces where labour is exchanged as a commodity. The firms that operate them consider workers as “independent contractors” without the social and economic benefits of traditional employment relations. This presentation explores the personal networks of Latin American data workers who train and verify data for machine learning algorithms from their homes. A series of in-depth interviews and an analysis of a self-completion questionnaire and web traffic data suggests that these workers are embedded of networks of trusts build on online and offline interactions. These findings show a continuation of exploitative supply chains in the current artificial intelligence market, where wealthy companies and research institutions in advanced economies profit from the economic and political situation of developing countries to access disembedded labor. This paper concludes by arguing that, though outsourced online labour, artificial intelligence developers not only extract value from their workers, but also indirectly from their communities and personal networks.

Predicting carcass cut yields in cattle from digital images using artificial intelligence

Deep Learning (DL) has proven to be a successful tool for many image classification problems but has yet to be applied to carcass images. The aim of this study was to train DL models to predict carcass cut yields and compare predictions to more standard machine learning (ML) methods. Three approaches were undertaken to predict the grouped carcass cut yields of Grilling cuts and Roasting cuts from a large dataset of 54,598 and 69,246 animals respectively. The approaches taken were (1) animal phenotypic data used as features for a range of ML algorithms, (2) carcass images used to train Convolutional Neural Networks, and (3) carcass dimensions measured directly from the carcass images, combined with the associated phenotypic data and used as feature data for ML algorithms. Results showed that DL models can be trained to predict carcass cuts yields but an approach that uses carcass dimensions in ML algorithms performs slightly better in absolute terms.

DETECTION AND CLASSIFICATION OF SYMBOLS IN PRINCIPLE SKETCHES USING DEEP LEARNING

AbstractData-driven methods from the field of Artificial Intelligence or Machine Learning are increasingly applied in mechanical engineering. This refers to the development of digital engineering in recent years, which aims to bring these methods into practice in order to realize cost and time savings. However, a necessary step towards the implementation of such methods is the utilization of existing data. This problem is essential because the mere availability of data does not automatically imply data usability. Therefore, this paper presents a method to automatically recognize symbols from principle sketches, which allows the generation of training data for machine learning algorithms. In this approach, the symbols are created randomly and their illustration varies with each generation. . A deep learning network from the field of computer vision is used to test the generated data set and thus to recognize symbols on principle sketches. This type of drawing is especially interesting because the cost-saving potential is very high due to the application in the early phases of the product development process.

Digital twin-driven machine learning: ball bearings fault severity classification

Machine learning algorithms (MLAs) are increasingly being used as effective techniques for processing vibration signals obtained from complex industrial machineries. Previous applications of automatic fault detection algorithms in the diagnosis of rotating machines were mainly based on historical operating data sets, limiting the diagnostic reliability to devices with an extended operating history. Moreover, physically collected data are often restricted by the conditions of acquisition and the specific elements for which they were recorded. Digital twin (DT) provides a powerful tool able to generate a huge amount of training data for MLAs. However, the DT model must be accurate enough to substitute the experiments. This work aims to escape the experience requirement by using a simulation-driven MLA based on the multifactorial analysis of fault indicators associated with a DT. To achieve this approach, a numerical model of a rotor-ball bearing system is developed. The latter is updated according to a parameter update scheme based on a comparison between the relevant features of the experimentally measured signals and the signals simulated by the model. These features are chosen as the selected input parameters of the MLA classifier. The results show that after updating, the developed DT has provided a reliable diagnostic with an adaptive degradation analysis, which makes the simulated data suitable for the construction of a machine learning predictive model. Two common MLAs, (multi-kernel support vector machine) and (k nearest neighbor’s algorithm), were trained using the simulated data and validated later against experimental datasets.

Inrush Current Measurement for Transient Space Characterization and Fault Detection

Training the load identification algorithm for a power-system monitor (PSM) begins with characterization or observation of the loads. This article demonstrates a phase-controlled switch that, in situ, accounts for the effects of source and line impedance on load transients at controllable phase angles with respect to the ac voltage line-cycle. The phase angle significantly affects the inrush for many loads, such as lighting, induction motors, and power supplies. The circuitry and signal identification techniques presented in this article can provide a full set of training data for machine learning algorithms for power system monitoring. The system can also be used as the foundation of a “smart switch” in grid applications for automatic demand response control.

Enabling Remote Patient Monitoring Through the Use of Smart Thermostat Data in Canada: Exploratory Study.

BackgroundAdvances in technology have made the development of remote patient monitoring possible in recent years. However, there is still room for innovation in the types of technologies that are developed, used, and implemented. The smart thermostat solutions provided in this study can expand beyond typically defined features and be used for improved holistic health monitoring purposes.ObjectiveThe aim of this study is to validate the hypothesis that remote motion sensors could be used to quantify and track an individual’s movements around the house. On the basis of our results, the next step would be to determine if using remote motion sensors could be a novel data collection method compared with the national census-level surveys administered by governmental bodies. The results will be used to inform a more extensive implementation study of similar smart home technologies to gather data for machine learning algorithms and to build upon pattern recognition and comprehensive health monitoring.MethodsWe conducted a pilot study with a sample size of 8 to validate the use of remote motion sensors to quantify movement in the house. A large database containing data from smart home thermostats was analyzed to compare the following indicators; sleep, physical activity, and sedentary behavior. These indicators were developed by the Public Health Agency of Canada and are collected through traditional survey methods.ResultsThe results showed a significant Spearman rank correlation coefficient of 0.8 (P<.001), which indicates a positive linear association between the total number of sensors activated and the total number of indoor steps traveled by study participants. In addition, the indicators of sleep, physical activity, and sedentary behavior were all found to be highly comparable with those attained by the Public Health Agency of Canada.ConclusionsThe findings demonstrate that remote motion sensors data from a smart thermostat solution are a viable option when compared with traditional survey data collection methods for health data collection and are also a form of zero-effort technology that can be used to monitor the activity levels and nature of activity of occupants within the home.

Quantification of Interpatient 12-lead ECG Variabilities within a Healthy Cohort

Abstract The morphology of the electrocardiogram (ECG) varies among different healthy subjects due to anatomical and structural reasons, such as for example the shape of the heart geometry or the position and size of surrounding organs in the torso. Knowledge about these ECG morphology changes could be used to parameterize electrophysiological simulations of the human heart. In this work, we detected the boundaries of ECG waveforms, i.e. the P-wave, the QRS-complex and the T-wave, in 12- lead ECGs from 918 healthy subjects in the Physionet Computing in Cardiology Challenge 2020 Database with the IBT openECG toolbox. Subsequently, we obtained the onset, the peak and the offset of each P-wave, QRS-complex and T-wave in the signal. In this way, the duration of the P-wave, the QRScomplex and the T-wave, the PQ-, RR- and the QT-interval as well as the amplitudes of the P-wave, the Q-, R- and Speak and the T-wave in each lead were extracted from the 918 healthy ECGs. Their statistical distributions and correlation between each other were assessed. The highest variabilities among the 918 healthy subject were found for the RR interval and the amplitudes of the QRScomplex. The highest correlation was observed for feature pairs that represent the same feature in different leads. Especially the R-peak amplitudes showed a strong correlation across different leads. The calculated feature distributions can be used to optimize the parameters of populations of cardiac electrophysiological models. In this way, realistic in-silico generated surface ECGs can be simulated in large scale and could be used as input data for machine learning algorithms for a classification of cardiovascular diseases.