Train Machine Learning Models Research Articles

Tumor detection by analysis of both symmetric- and hemi-methylation of plasma cell-free DNA

Aberrant DNA methylation patterns have been used for cancer detection. However, DNA hemi-methylation, present at about 10% CpG dinucleotides, has been less well studied. Here we show that a majority of differentially hemi-methylated regions (DHMRs) in liver tumor DNA or plasma cells free (cf) DNA do not overlap with differentially methylated regions (DMRs) of the same samples, indicating that DHMRs could serve as independent biomarkers. Furthermore, we analyzed the cfDNA methylomes of 215 samples from individuals with liver or brain cancer and individuals without cancer (controls), and trained machine learning models using DMRs, DHMRs or both. The models incorporated with both DMRs and DHMRs show a superior performance compared to models trained with DMRs or DHMRs, with AUROC being 0.978, 0.990, and 0.983 in distinguishing control, liver and brain cancer, respectively, in a validation cohort. This study supports the potential of utilizing both DMRs and DHMRs for multi-cancer detection.

Designing of potential polymer donors with lower reorganization energy and chemical space exploration in search of materials for efficient organic solar cells

Machine learning with fascinating material predicting capabilities has emerged as a growing throughput artificial intelligence technology to accelerate the design, screening and prediction of material properties. In this study, chemical descriptors are employed to train machine learning (ML) models for the designing of potential polymer donor materials aimed at enhancing the efficiency of organic solar cells (OSCs). To achieve this, ten ML models are evaluated. Among all, the random forest regressor model has shown better predictive ability. In addition, to quantify the importance of features, the numerical scores of the significance of the features are determined through Feature importance analysis. Moreover, 10k new polymers are generated employing the python-based method and their reorganization energies are further predicted using a pre-trained ML model. The assemblage of polymers is done on the basis of the reorganization energy, those with lower reorganization energy being preferred. Furthermore, the synthetic accessibility and thus the structural diversity among the chosen polymers donors are examined, revealing a structural difference among the chosen polymers. This work paves a way for the designing and screening of efficient semiconductors with low reorganization energy to increase the efficiency of OSCs.

Prediction of esports competition outcomes using EEG data from expert players

Considerable efforts have focused on predicting sports event outcomes to enhance spectating, coaching, and betting. However, accurately predicting match results across various competitive scenes is challenging. Many researchers have achieved reasonable prediction accuracy using machine learning (ML) models and static data, such as player statistics up to the last match, to perform binary win/loss classification. However, the inherent uncertainty of sports often leads to unexpected outcomes. For example, narrowing the skill-level gap between players may negatively impact prediction accuracy. Moreover, accurately predicting “upsets,” where lower-skilled players defeat higher-skilled opponents, remains challenging. Conventional win/loss prediction techniques rely on static attributes derived from past match statistics, which fail to capture the player's condition immediately before the match. To address this limitation, this study focused on dynamic information reflecting the pre-match player's condition, particularly electroencephalography (EEG) data, recognized as a potent mental conditioning biomarker. We collected pre-match EEG data from esports experts and trained ML models on these data, aiming to assess the ability of ML techniques trained on EEG data to predict match outcomes, particularly in cases subject to high unpredictability. The findings revealed that the light gradient boosting machine (LightGBM) algorithm trained on EEG data achieved the highest prediction accuracy (80%), with parietal beta activity being the most relevant feature. Furthermore, predictive performance remained consistent even in match scenarios involving similar-level players and upset situations. Hence, this approach extends the applicability of win/loss prediction to traditionally unpredictable sports scenes, enhancing the quality of spectator experiences.

TelescopeML – I. An End-to-End Python Package for Interpreting Telescope Datasets through Training Machine Learning Models, Generating Statistical Reports, and Visualizing Results

TelescopeML – I. An End-to-End Python Package for Interpreting Telescope Datasets through Training Machine Learning Models, Generating Statistical Reports, and Visualizing Results

Optimizing data acquisition: a Bayesian approach for efficient machine learning model training

Acquiring a substantial number of data points for training accurate machine learning (ML) models is a big challenge in scientific fields where data collection is resource-intensive. Here, we propose a novel approach for constructing a minimal yet highly informative database for training ML models in complex multi-dimensional parameter spaces. To achieve this, we mimic the underlying relation between the output and input parameters using Gaussian process regression (GPR). Using a set of known data, GPR provides predictive means and standard deviation for the unknown data. Given the predicted standard deviation by GPR, we select data points using Bayesian optimization to obtain an efficient database for training ML models. We compare the performance of ML models trained on databases obtained through this method, with databases obtained using traditional approaches. Our results demonstrate that the ML models trained on the database obtained using Bayesian optimization approach consistently outperform the other two databases, achieving high accuracy with a significantly smaller number of data points. Our work contributes to the resource-efficient collection of data in high-dimensional complex parameter spaces, to achieve high precision ML predictions.

Estimation of Water Velocity and Receiver Position Changes in Time-Lapse Seismic Using Machine Learning

Non-repeatability is one of the main obstacles in time-lapse seismic, as it significantly degrades the interpretation of reservoir-related signals. Correcting the data variations caused by non-repeatability (4D noise) is of paramount importance, which usually requires the estimation of the changing parameter. ln this paper, we propose a Machine Learning (ML) workflow for the quantitative estimation of two types of 4D noise: changes in the speed of sound in water and receiver lateral positions. A synthetic database, modeled from a velocity model estimated from the Brazilian pre-salt, and containing many time-lapse seismic surveys was generated for the supervised training of ML models. lnput samples consist of subsets of common-shot seismograms. We studied many combinations of ML regression algorithms and feature extraction techniques, for scenarios where data is contaminated, or not, with Gaussian random noise. The regression algorithms considered were Fully-Connected Neural Network, Extreme Gradient Boosting, and Bayesian Ridge. Four rectangular crops of the common-shot seismogram were tested as input features: full seismogram, first half of time samples, 11 smallest-offset traces, and the time samples focusing on the first arrivals region. The combination with the best trade-off between accuracy and model complexity is the Bayesian Ridge fed with the 11 smallest-offset traces, which estimated position and velocity time-lapse changes with median accuracy of 0.115 m and 0.017 m/s for the case with Gaussian noise. Besides the correction of repeatability-related variations, our results are useful in the 4D Full-Waveform lnversion which needs accurate parameters to produce good seismic images.

An Oversampling Technique with Descriptive Statistics

Oversampling is often applied as a means to win a better knowledge model. Several oversampling methods based on synthetic instances have been suggested, and SMOTE is one of the representative oversampling methods that can generate synthetic instances of a minor class. Until now, the oversampled data has been used conventionally to train machine learning models without statistical analysis, so it is not certain that the machine learning models will be fine for unseen cases in the future. However, because such synthetic data is different from the original data, we may wonder how much it resembles the original data so that the oversampled data is worth using to train machine learning models. For this purpose, I conducted this study on a representative dataset called wine data in the UCI machine learning repository, which is one of the datasets that has been experimented with by many researchers in research for knowledge discovery models. I generated synthetic data iteratively using SMOTE, and I compared the synthetic data with the original data of wine to see if it was statistically reliable using a box plot and t-test. Moreover, since training a machine learning model by supplying more high-quality training instances increases the probability of obtaining a machine learning model with higher accuracy, it was also checked whether a better machine learning model of random forests can be obtained by generating much more synthetic data than the original data and using it for training the random forests. The results of the experiment showed that small-scale oversampling produced synthetic data with statistical characteristics that were statistically slightly different from the original data, but when the oversampling rate was relatively high, it was possible to generate data with statistical characteristics similar to the original data, in other words, after generating high-quality training data, and by using it to train the random forests, it was possible to generate random forests with higher accuracy than using the original data alone, from 97.75% to 100%. Therefore, by supplying additional statistically reliable synthetic data as a way of oversampling, it was possible to create a machine-learning model with a higher predictive rate.

Machine learning approach to evaluating impact behavior in fabric-laminated composite materials

Fabric-layered composites play a crucial role in safety and surveillance applications, making it imperative to accurately predict their impact behavior. This research focuses on creating a machine-learning model to predict the impact behavior of fabric-stacked composites, specifically carbon and Kevlar fabrics. Low-velocity impact tests were performed with varying parameters, and the impact energy and laminate thickness information were used to train machine-learning models to predict impact properties such as impact force, displacement, and absorbed energy. It was observed that the impact force increased by 118.5 % in carbon-laminated fibers and 175.8 % in Kevlar-laminated fibers, while the hybrid layer showed a 101.4 % increase upon impact from 16J. Displacement can affect the stability of the layered structure; thus, a similar stacking sequence is less stable than the hybrid laminated structure. In terms of absorbed energy, as the laminated layers increase, carbon fiber laminate absorbs 4.8 times more energy, and Kevlar fiber and hybrid laminated structures absorb 3 times more energy at higher impact energy. Furthermore, four machine learning models were used to investigate the impact behavior of identical and mixed-layered laminated composites. The impact force and displacement were predicted with higher accuracy using the polynomial regression model, achieving 80 % and 89 % accuracy, respectively. The support vector machine predicted the absorbed energy with approximately 96 % accuracy. In continuing, the experimental results closely matched the predictions made by the other machine learning models utilized in this study. Additionally, the importance of distinctive features and their influence on the performance of the machine learning-based model were interpreted using transposed features of importance and dependency plots. Various failure modes in fabric laminated composites were also identified, providing insights to enhance the performance of stacked fiber materials.

Statistical analysis of microstructurally small fatigue crack growth in three-dimensional polycrystals based on high-fidelity numerical simulations

Microstructurally small cracks (MSCs) are sensitive to their local microstructural neighborhoods, resulting in highly variable 3D crack propagation within individual samples and among a population of samples. Statistically quantifying this complex spatiotemporal variability requires collecting and analyzing a large number of 3D MSC growth observations, which remains challenging due to experimental constraints that limit the availability of 3D observations of MSCs in the existing literature. We address this gap by leveraging virtual observations of MSC growth using a state-of-the-art, high-fidelity simulation framework to provide a statistical perspective on 3D MSC growth behavior. The MSC growth simulations were performed in 40 unique but statistically similar polycrystalline microstructures, and the data extracted from the virtual observations were collectively analyzed to quantify variability in geometric, microstructural, and micromechanical aspects of MSCs. Variability in tortuosity, crack surface crystallography, and MSC growth parameters (local crack extension and kink angle) were quantified statistically. The results show that 3D MSC surfaces do not necessarily propagate along {111} crystallographic planes despite crack traces on radial planes being closely aligned with {111} slip plane traces. Also, the 3D MSC surfaces exhibit increasingly tortuous propagation and spatially varying local crack growth rates (with variability as high as 59% among the population). Moreover, a correlation study revealed some of the highly influential microstructural and micromechanical features controlling MSC growth parameters. The quantified variability holds direct implications for advanced materials prognosis in engineering applications, and the identified features from the correlation study can be leveraged to train machine learning models for rapidly predicting MSC growth behavior in the future.

Machine Learning Model Training Framework for Nonlinear Signal-to-Noise Ratio Estimation in Heterogeneous Optical Networks

Machine Learning Model Training Framework for Nonlinear Signal-to-Noise Ratio Estimation in Heterogeneous Optical Networks

Generating Synthetic MR Spectroscopic Imaging Data with Generative Adversarial Networks to Train Machine Learning Models.

To develop a new method to generate synthetic MR spectroscopic imaging (MRSI) data for training machine learning models. This study targeted routine MRI examination protocols with single voxel spectroscopy (SVS). A novel model derived from pix2pix generative adversarial networks was proposed to generate synthetic MRSI data using MRI and SVS data as inputs. T1- and T2-weighted, SVS, and reference MRSI data were acquired from healthy brains with clinically available sequences. The proposed model was trained to generate synthetic MRSI data. Quantitative evaluation involved the calculation of the mean squared error (MSE) against the reference and metabolite ratio value. The effect of the location of and the number of the SVS data on the quality of the synthetic MRSI data was investigated using the MSE. The synthetic MRSI data generated from the proposed model were visually closer to the reference. The 95% confidence interval (CI) of the metabolite ratio value of synthetic MRSI data overlapped with the reference for seven of eight metabolite ratios. The MSEs tended to be lower in the same location than in different locations. The MSEs among groups of numbers of SVS data were not significantly different. A new method was developed to generate MRSI data by integrating MRI and SVS data. Our method can potentially increase the volume of MRSI data training for other machine learning models by adding SVS acquisition to routine MRI examinations.

Machine learning assisted designing of polymers and refractive index prediction: Easy and fast screening of polymers from chemical space

For the selection of efficient materials capable of optical applicability, the property of refractive index (RI) of the material is considered extremely significant. Acquisition of RI values via empirical framework is difficult and prolonged task. Consequently, data-driven method provides a rapid alternative for estimating RI values. This study presents a fast framework that is based on machine learning (ML) approach to design novel polymers capable of performing in optical applications. The framework involves training of ML models to predict the RI values of polymers. Subsequently, 10,000 new polymers were generated utilizing the Breaking Retrosynthetically Interesting Chemical Substructures (BRICS) methodology and fast ML model is used to predict the RI values of newly generated polymers. Polymers exhibiting higher RI values were retained. The synthetic accessibility of these selected polymers was also assessed in order to facilitate the future empirical measurements. Furthermore, chemical similarity among the chosen polymers was also investigated and structural diversity was revealed among the selected polymers. This study is introducing fast and easy framework for the designing and screening of efficient materials. This framework can be modified to study other materials and properties.

Progressive transfer learning for advancing machine learning-based reduced-order modeling

To maximize knowledge transfer and improve the data requirement for data-driven machine learning (ML) modeling, a progressive transfer learning for reduced-order modeling (p-ROM) framework is proposed. A key concept of p-ROM is to selectively transfer knowledge from previously trained ML models and effectively develop a new ML model(s) for unseen tasks by optimizing information gates in hidden layers. The p-ROM framework is designed to work with any type of data-driven ROMs. For demonstration purposes, we evaluate the p-ROM with specific Barlow Twins ROMs (p-BT-ROMs) to highlight how progress learning can apply to multiple topological and physical problems with an emphasis on a small training set regime. The proposed p-BT-ROM framework has been tested using multiple examples, including transport, flow, and solid mechanics, to illustrate the importance of progressive knowledge transfer and its impact on model accuracy with reduced training samples. In both similar and different topologies, p-BT-ROM achieves improved model accuracy with much less training data. For instance, p-BT-ROM with four-parent (i.e., pre-trained models) outperforms the no-parent counterpart trained on data nine times larger. The p-ROM framework is poised to significantly enhance the capabilities of ML-based ROM approaches for scientific and engineering applications by mitigating data scarcity through progressively transferring knowledge.

Contrastive Self-supervised Learning for Neurodegenerative Disorder Classification.

Neurodegenerative diseases such as Alzheimer's disease (AD) or frontotemporal lobar degeneration (FTLD) involve specific loss of brain volume, detectable in vivo using T1-weighted MRI scans. Supervised machine learning approaches classifying neurodegenerative diseases require diagnostic-labels for each sample. However, it can be difficult to obtain expert labels for a large amount of data. Self-supervised learning (SSL) offers an alternative for training machine learning models without data-labels. We investigated if the SSL models can applied to distinguish between different neurodegenerative disorders in an interpretable manner. Our method comprises a feature extractor and a downstream classification head. A deep convolutional neural network trained in a contrastive self-supervised way serves as the feature extractor, learning latent representation, while the classifier head is a single-layer perceptron. We used N=2694 T1-weighted MRI scans from four data cohorts: two ADNI datasets, AIBL and FTLDNI, including cognitively normal controls (CN), cases with prodromal and clinical AD, as well as FTLD cases differentiated into its sub-types. Our results showed that the feature extractor trained in a self-supervised way provides generalizable and robust representations for the downstream classification. For AD vs. CN, our model achieves 82% balanced accuracy on the test subset and 80% on an independent holdout dataset. Similarly, the Behavioral variant of frontotemporal dementia (BV) vs. CN model attains an 88% balanced accuracy on the test subset. The average feature attribution heatmaps obtained by the Integrated Gradient method highlighted hallmark regions, i.e., temporal gray matter atrophy for AD, and insular atrophy for BV. In conclusion, our models perform comparably to state-of-the-art supervised deep learning approaches. This suggests that the SSL methodology can successfully make use of unannotated neuroimaging datasets as training data while remaining robust and interpretable.

Bayesian Networks for Prescreening in Depression: Algorithm Development and Validation.

Identifying individuals with depressive symptomatology (DS) promptly and effectively is of paramount importance for providing timely treatment. Machine learning models have shown promise in this area; however, studies often fall short in demonstrating the practical benefits of using these models and fail to provide tangible real-world applications. This study aims to establish a novel methodology for identifying individuals likely to exhibit DS, identify the most influential features in a more explainable way via probabilistic measures, and propose tools that can be used in real-world applications. The study used 3 data sets: PROACTIVE, the Brazilian National Health Survey (PesquisaNacionalde Saúde [PNS]) 2013, and PNS 2019, comprising sociodemographic and health-related features. A Bayesian network was used for feature selection. Selected features were then used to train machine learning models to predict DS, operationalized as a score of ≥10 on the 9-item Patient Health Questionnaire. The study also analyzed the impact of varying sensitivity rates on the reduction of screening interviews compared to a random approach. The methodology allows the users to make an informed trade-off among sensitivity, specificity, and a reduction in the number of interviews. At the thresholds of 0.444, 0.412, and 0.472, determined by maximizing the Youden index, the models achieved sensitivities of 0.717, 0.741, and 0.718, and specificities of 0.644, 0.737, and 0.766 for PROACTIVE, PNS 2013, and PNS 2019, respectively. The area under the receiver operating characteristic curve was 0.736, 0.801, and 0.809 for these 3 data sets, respectively. For the PROACTIVE data set, the most influential features identified were postural balance, shortness of breath, and how old people feel they are. In the PNS 2013 data set, the features were the ability to do usual activities, chest pain, sleep problems, and chronic back problems. The PNS 2019 data set shared 3 of the most influential features with the PNS 2013 data set. However, the difference was the replacement of chronic back problems with verbal abuse. It is important to note that the features contained in the PNS data sets differ from those found in the PROACTIVE data set. An empirical analysis demonstrated that using the proposed model led to a potential reduction in screening interviews of up to 52% while maintaining a sensitivity of 0.80. This study developed a novel methodology for identifying individuals with DS, demonstrating the utility of using Bayesian networks to identify the most significant features. Moreover, this approach has the potential to substantially reduce the number of screening interviews while maintaining high sensitivity, thereby facilitating improved early identification and intervention strategies for individuals experiencing DS.

O-131 A novel approach to predicting the potential of a blastocyst using untargeted spectral analysis of spent culture medium and machine learning

Abstract Study question Is it possible to effectively predict negative outcomes of intrauterine embryo transfer using modified untargeted Raman spectroscopy of the spent culture medium? Summary answer The proposed spectroscopic approach, combined with machine learning methods, allows identifying a group of embryos that are almost unable to implant and develop ongoing pregnancy. What is known already Embryo deselection is usually based on the karyotype analysis of the embryo and originates from the assumption that an abnormal PGT result is associated with negative transfer outcomes. Embryo selection methods aimed at identifying the most capable for implantation embryo are based on morphological, morphometric, and metabolic approaches. However, due to the dependence of the phenomenon (pregnancy) on several factors, predicting the outcome of the transfer solely through variables derived from the embryo does not correspond to modern views on implantation. At the same time, predicting negative transfer outcomes through the information originating from the embryo appears quite rational. Study design, size, duration 424 samples of spent culture medium used, collected from microdrops with individually cultured blastocyst from the internal institutional biosamples repository. Medium samples were collected from July 1, 2021, to June 1, 2023. The class attribute for each sample was the absence or presence of ongoing pregnancy at the 12-14 week of gestational age. The positive class (ongoing pregnancy) proportion in the presented dataset is 33%, and the negative (no ongoing pregnancy) class proportion is 67%. Participants/materials, setting, methods The study material was spent culture medium from blastocysts with known transfer outcomes by the end of first trimester. All media samples were stored at -86 °С until the moment of spectroscopy. Before the study, samples were thawed at room temperature for 15 minutes. Untargeted Raman spectroscopy was done on a signal amplification substrate using a green laser (532 nm). The architecture selection and machine learning model training were performed using the Python library mcfly. Main results and the role of chance After building the model the following results for validation and test sets were obtained: So it is possible to identify a group of embryos that are almost incapable of producing ongoing pregnancy. This group constitutes about 40% of the total number of negative outcomes. The error of determining an embryo as being incapable for implantation and normal pregnancy development is about 6%, what is relatively low number. Limitations, reasons for caution The presented data are not multicenter; they were obtained from samples from one laboratory and using a limited set (only one manufacturer) of culture media. This may lead to reduced or nonexistent relevance of the model for data from other sources and conditions. Wider implications of the findings It is necessary to check the proposed method on an additional set of culture media from other manufactures and laboratories. This may lead to refining the parameters of the built neural network. The assumption of significant clinical value of the developed model needs to be checked through prospective testing. Trial registration number fpc-2023-ivf-01

Integrating Machine Learning and Feature Extraction for Islanding Detection in Grid-Connected Photovoltaic Systems: A Hybrid Intelligent Approacha

Introduction: In recent years, the integration of renewable energy sources, particularly Photovoltaic (PV) systems, into the grid has garnered considerable attention. However, the distributed nature of these grid-integrated PV systems has introduced challenges concerning grid faults and maintenance. Method: This paper aims to present a pioneering approach to augment the monitoring of gridintegrated PV systems by integrating intelligent methods, specifically machine learning and feature extraction techniques. The primary focus of this approach is on islanding detection, which involves promptly identifying grid faults or maintenance challenges and initiating the grid's transition to an isolated mode of operation. To accomplish this, an intelligent signaling method is employed, capitalizing on the capabilities of Distributed Generation (DG) networks. By recording critical signals such as voltage, current, and frequency at common coupling points, fault conditions can be accurately detected. Signals obtained from the grid are subjected to wavelet transformation to extract pertinent information that characterizes fault conditions. These extracted features are then utilized as inputs to machine learning methods, facilitating the proposal of intelligent islanding scenarios. To assess the efficacy of the proposed approach, simulations are conducted on a grid-connected PV system. The recorded signals are meticulously analyzed, and the extracted features are employed to train machine learning models. The performance of these models is evaluated based on their ability to detect fault conditions and initiate appropriate islanding scenarios accurately. Result: The results obtained demonstrate the immense potential of the proposed approach in bolstering the monitoring of grid-integrated PV systems. Conclusion: By synergizing machine learning techniques with feature extraction and intelligent signaling methods with the KNN-Confusion Matrix, all predicted labels match the true labels, resulting in a 100% accuracy. The detection of grid faults and maintenance challenges can be substantially improved, thereby fostering more efficient and dependable operation of these systems.

Detection of disease-specific signatures in B cell repertoires of lymphomas using machine learning.

The classification of B cell lymphomas-mainly based on light microscopy evaluation by a pathologist-requires many years of training. Since the B cell receptor (BCR) of the lymphoma clonotype and the microenvironmental immune architecture are important features discriminating different lymphoma subsets, we asked whether BCR repertoire next-generation sequencing (NGS) of lymphoma-infiltrated tissues in conjunction with machine learning algorithms could have diagnostic utility in the subclassification of these cancers. We trained a random forest and a linear classifier via logistic regression based on patterns of clonal distribution, VDJ gene usage and physico-chemical properties of the top-n most frequently represented clonotypes in the BCR repertoires of 620 paradigmatic lymphoma samples-nodular lymphocyte predominant B cell lymphoma (NLPBL), diffuse large B cell lymphoma (DLBCL) and chronic lymphocytic leukemia (CLL)-alongside with 291 control samples. With regard to DLBCL and CLL, the models demonstrated optimal performance when utilizing only the most prevalent clonotype for classification, while in NLPBL-that has a dominant background of non-malignant bystander cells-a broader array of clonotypes enhanced model accuracy. Surprisingly, the straightforward logistic regression model performed best in this seemingly complex classification problem, suggesting linear separability in our chosen dimensions. It achieved a weighted F1-score of 0.84 on a test cohort including 125 samples from all three lymphoma entities and 58 samples from healthy individuals. Together, we provide proof-of-concept that at least the 3 studied lymphoma entities can be differentiated from each other using BCR repertoire NGS on lymphoma-infiltrated tissues by a trained machine learning model.

A probabilistic approach to training machine learning models using noisy data

Machine learning (ML) models are increasingly popular in environmental and hydrologic modeling, but they typically contain uncertainties resulting from noisy data (erroneous or outlier data). This paper presents a novel probabilistic approach that combines ML and Markov Chain Monte Carlo simulation to (1) detect and underweight likely noisy data, (2) develop an approach capable of detecting noisy data during model deployment, and (3) interpret the reasons why a data point is deemed noisy to help heuristically distinguish between outliers and erroneous data. The new algorithm recognizes that there is no unique way to split the training data into noisy and clean data, and thus produces an ensemble of plausible splits. The algorithm successfully detected noisy data in synthetic benchmark problems with varying complexity and a real-world public supply water withdrawal dataset. The algorithm is generic and flexible, making it suitable for application across a broad range of hydrologic and environmental disciplines.

Hawk: Accurate and Fast Privacy-Preserving Machine Learning Using Secure Lookup Table Computation

Training machine learning models on data from multiple entities without direct data sharing can unlock applications otherwise hindered by business, legal, or ethical constraints. In this work, we design and implement new privacy-preserving machine learning protocols for logistic regression and neural network models. We adopt a two-server model where data owners secret-share their data between two servers that train and evaluate the model on the joint data. A significant source of inefficiency and inaccuracy in existing methods arises from using Yao’s garbled circuits to compute non-linear activation functions. We propose new methods for computing non-linear functions based on secret-shared lookup tables, offering both computational efficiency and improved accuracy. Beyond introducing leakage-free techniques, we initiate the exploration of relaxed security measures for privacy-preserving machine learning. Instead of claiming that the servers gain no knowledge during the computation, we contend that while some information is revealed about access patterns to lookup tables, it maintains epsilon-dX-privacy. Leveraging this relaxation significantly reduces the computational resources needed for training. We present new cryptographic protocols tailored to this relaxed security paradigm and define and analyze the leakage. Our evaluations show that our logistic regression protocol is up to 9x faster, and the neural network training is up to 688x faster than SecureML. Notably, our neural network achieves an accuracy of 96.6% on MNIST in 15 epochs, outperforming prior benchmarks that capped at 93.4% using the same architecture.