Representative Training Dataset Research Articles

COMPUTED TOMOGRAPHIC ANGIOGRAPHY DATA SET CONTAINING ABDOMINAL AORTIC SEGMENTATION

BACKGROUND: Artificial Intelligence (AI) is widely actively in radiology. The effectiveness of AI algorithms essentially depends on the availability of relevant and representative training and test datasets. Currently, such datasets are highly demanded in open access, particularly CT angiographic (CTA) studies of the abdominal aorta containing not only pathology classification but also vessel segmentation. The disadvantages of existing solutions include small sample volume, narrow specialization, and disparate data preparation methodology. AIM – Design of a CTA dataset containing abdominal aortic segmentation for cases of normal, dilatation, thrombosis, and calcinosis. METHODS: In accordance with the methodology of AI-algorithms testing, the terms of reference for datasets preparation were developed, the necessary sample volume was calculated and approval of the independent ethical committee was obtained. Previously designed original algorithm of semi-automatic segmentation of abdominal aorta on CTA-studies using open access software Slicer 3D was used. Inclusion criteria: CTA studies, or abdominal CT studies with contrast enhancement; presence of arterial phase of the scan; slice thickness not more than 3 mm. Exclusion criteria: presence of any foreign objects in the aortic lumen, aortic dissection. The algorithm was tested on patient data obtained from the Unified Radiologic Information System. Expert evaluation of the compliance of the obtained results with the formed requirements, as well as the assessment of time costs when using the developed segmentation algorithm was carried out. RESULTS: The calculated sample size was 100 CTA studies with a slice thickness ≤ 1.2 mm. Population data: the number of unique patients was 100, the proportion of female patients was 51%; the median age was 62 years ranged from 18 to 84 years. The proportion of pathologic studies was 61%, of which (including combined pathology): 60 cases contained calcification, both by 18 cases with aortic dilatation and thrombosed lumen. Different non-overlapping masks were used for each object. The dataset is presented in the public domain. The average processing time is 0.8 hours per 100 CT-images. CONCLUSION: The public dataset containing 100 CTA studies of segmented abdominal aorta for normal, dilated, thrombotic and calcified cases was created. Results are valuable for AI-algorithms design and anthropomorphic 3D-modeling.

Representative training data sets are critical for accurate machine-learning classification of microscopy images of particles formed by lipase-catalyzed polysorbate hydrolysis.

Polysorbate 20 (PS20) is commonly used as an excipient in therapeutic protein formulations. However, over the course of a therapeutic protein product's shelf life, minute amounts of co-purified host-cell lipases may cause slow hydrolysis of PS20, releasing fatty acids (FAs). These FAs may precipitate to form subvisible particles that can be detected and imaged by various techniques, e.g., flow imaging microscopy (FIM). Images of particles can then be classified using supervised convolutional neural networks (CNNs). However, CNNs should be trained on representative images of particles which, as we demonstrate in this work, may be challenging to obtain. Here, we tested several rapid techniques to create FA particles and examined whether CNNs trained on microscopy images of these rapidly formed particles could accurately classify images of particles that had been produced by kinetically slower lipase-catalyzed hydrolysis of PS20. CNNs trained on images of rapidly produced particles were less accurate in classifying images of FA particles that had been produced by enzymatic hydrolysis of PS20 than CNNs trained with images of particles generated by the same slow hydrolysis, highlighting the importance of using representative image data sets for training CNN classifiers.

Crash testing machine learning force fields for molecules, materials, and interfaces: molecular dynamics in the TEA challenge 2023.

We present the second part of the rigorous evaluation of modern machine learning force fields (MLFFs) within the TEA Challenge 2023. This study provides an in-depth analysis of the performance of MACE, SO3krates, sGDML, SOAP/GAP, and FCHL19* in modeling molecules, molecule-surface interfaces, and periodic materials. We compare observables obtained from molecular dynamics (MD) simulations using different MLFFs under identical conditions. Where applicable, density-functional theory (DFT) or experiment serves as a reference to reliably assess the performance of the ML models. In the absence of DFT benchmarks, we conduct a comparative analysis based on results from various MLFF architectures. Our findings indicate that, at the current stage of MLFF development, the choice of ML model is in the hands of the practitioner. When a problem falls within the scope of a given MLFF architecture, the resulting simulations exhibit weak dependency on the specific architecture used. Instead, emphasis should be placed on developing complete, reliable, and representative training datasets. Nonetheless, long-range noncovalent interactions remain challenging for all MLFF models, necessitating special caution in simulations of physical systems where such interactions are prominent, such as molecule-surface interfaces. The findings presented here reflect the state of MLFF models as of October 2023.

Automatic similarity feature selection for ontology matching with semantic sampling

Ontology provides a formal representation of knowledge as a set of concepts within a domain and the relationships between them. Ontology Matching (OM) identifies similar entities across different ontologies. Similarity Features (SFs) characterize the similarity between two entities from different perspectives, which serves as the foundation of OM method. However, noisy and redundant SFs can obstruct the relevance of useful ones, reducing the quality of matching results. To offer a practical and efficient SF selection, this work proposes a new automatic SF selection framework for OM, which consists of two new components. First, a semantic sampling method is proposed to automatically construct a balanced and representative training dataset, without the need for expert intervention. This method first generates a basic sample set that considers both the representativeness of the concepts and the heterogeneous characteristics of the two ontologies, and then over-samples the minority samples based on their semantic context. Second, a wrapper-based SF selection method is designed, which includes an Adaptive Compact Genetic Algorithm (ACGA) and a Relaxed Naive Bayes (RNB) classifier. The ACGA is able to efficiently identify the SF subset, reducing the computational cost by adaptively maintaining a population probability model based on the diversity and confidence of the SFs. Moreover, the RNB can effectively evaluate the quality of selected SFs by modifying the feature independence assumption with recombination and weighting strategies. Experimental results on the ontology alignment evaluation initiative’s Benchmark demonstrate that our method can efficiently find accurate OM results, and significantly outperform the state-of-the-art matching techniques.

Land Surface Longwave Radiation Retrieval from ASTER Clear-Sky Observations

Surface longwave radiation (SLR) plays a pivotal role in the Earth’s energy balance, influencing a range of environmental processes and climate dynamics. As the demand for high spatial resolution remote sensing products grows, there is an increasing need for accurate SLR retrieval with enhanced spatial detail. This study focuses on the development and validation of models to estimate SLR using measurements from the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) sensor. Given the limitations posed by fewer spectral bands and data products in ASTER compared to moderate-resolution sensors, the proposed approach combines an atmospheric radiative transfer model MODerate resolution atmospheric TRANsmission (MODTRAN) with the Light Gradient Boosting Machine algorithm to estimate SLR. The MODTRAN simulations were performed to construct a representative training dataset based on comprehensive global atmospheric profiles and surface emissivity spectra data. Global sensitivity analyses reveal that key inputs influencing the accuracy of SLR retrievals should reflect surface thermal radiative signals and near-surface atmospheric conditions. Validated against ground-based measurements, surface upward longwave radiation (SULR) and surface downward longwave radiation (SDLR) using ASTER thermal infrared bands and surface elevation estimations resulted in root mean square errors of 17.76 W/m2 and 25.36 W/m2, with biases of 3.42 W/m2 and 3.92 W/m2, respectively. Retrievals show systematic biases related to extreme temperature and moisture conditions, e.g., causing overestimation of SULR in hot humid conditions and underestimation of SDLR in arid conditions. While challenges persist, particularly in addressing atmospheric variables and cloud masking, this work lays a foundation for accurate SLR retrieval from high spatial resolution sensors like ASTER. The potential applications extend to upcoming satellite missions, such as the Landsat Next, and contribute to advancing high-resolution remote sensing capabilities for an improved understanding of Earth’s energy dynamics.

Optimising brain age estimation through transfer learning: A suite of pre-trained foundation models for improved performance and generalisability in a clinical setting.

Estimated age from brain MRI data has emerged as a promising biomarker of neurological health. However, the absence of large, diverse, and clinically representative training datasets, along with the complexity of managing heterogeneous MRI data, presents significant barriers to the development of accurate and generalisable models appropriate for clinical use. Here, we present a deep learning framework trained on routine clinical data (N up to 18,890, age range 18-96 years). We trained five separate models for accurate brain age prediction (all with mean absolute error ≤4.0 years, R2 ≥ .86) across five different MRI sequences (T2 -weighted, T2 -FLAIR, T1 -weighted, diffusion-weighted, and gradient-recalled echo T2 *-weighted). Our trained models offer dual functionality. First, they have the potential to be directly employed on clinical data. Second, they can be used as foundation models for further refinement to accommodate a range of other MRI sequences (and therefore a range of clinical scenarios which employ such sequences). This adaptation process, enabled by transfer learning, proved effective in our study across a range of MRI sequences and scan orientations, including those which differed considerably from the original training datasets. Crucially, our findings suggest that this approach remains viable even with limited data availability (as low as N = 25 for fine-tuning), thus broadening the application of brain age estimation to more diverse clinical contexts and patient populations. By making these models publicly available, we aim to provide the scientific community with a versatile toolkit, promoting further research in brain age prediction and related areas.

Investigation of Methods for Determining the Aspect Angle of a Spatially Distributed Target in its Radar Image

The article is devoted to the problem of determining the aspect angle of a surface ship based on its radar image obtained in synthetic aperture radars. Among the main methods of solving this problem are the classical Bayesian method of multi-alternative hypothesis testing and its modifications, and/or the method of classifying surface ships located at different angles using artificial neural networks (ANNs). The study demonstrates that achieving high efficiency in aspect recognition using ANNs requires significant computational resources, as well as an access large, representative, and scalable training dataset. ANNs demonstrates high performance in various observation conditions in case sufficient computational and time resources; however, it is noteworthy that their effective training requires a substantial amount of processing time, reaching several hours. At the same time, classical methods are capable of performing calculations in fractions of a second, even on relatively low-powered devices. It is also worth noting that as the number of recognised classes increases, ANNs may consume up to tens of gigabytes of RAM, limiting the accessibility of this method in the aspect of task recognition of spatially distributed targets.

Bidirectional recurrent learning of inverse dynamic models for robots with elastic joints: a real-time real-world implementation.

Collaborative robots, or cobots, are designed to work alongside humans and to alleviate their physical burdens, such as lifting heavy objects or performing tedious tasks. Ensuring the safety of human-robot interaction (HRI) is paramount for effective collaboration. To achieve this, it is essential to have a reliable dynamic model of the cobot that enables the implementation of torque control strategies. These strategies aim to achieve accurate motion while minimizing the amount of torque exerted by the robot. However, modeling the complex non-linear dynamics of cobots with elastic actuators poses a challenge for traditional analytical modeling techniques. Instead, cobot dynamic modeling needs to be learned through data-driven approaches, rather than analytical equation-driven modeling. In this study, we propose and evaluate three machine learning (ML) approaches based on bidirectional recurrent neural networks (BRNNs) for learning the inverse dynamic model of a cobot equipped with elastic actuators. We also provide our ML approaches with a representative training dataset of the cobot's joint positions, velocities, and corresponding torque values. The first ML approach uses a non-parametric configuration, while the other two implement semi-parametric configurations. All three ML approaches outperform the rigid-bodied dynamic model provided by the cobot's manufacturer in terms of torque precision while maintaining their generalization capabilities and real-time operation due to the optimized sample dataset size and network dimensions. Despite the similarity in torque estimation of these three configurations, the non-parametric configuration was specifically designed for worst-case scenarios where the robot dynamics are completely unknown. Finally, we validate the applicability of our ML approaches by integrating the worst-case non-parametric configuration as a controller within a feedforward loop. We verify the accuracy of the learned inverse dynamic model by comparing it to the actual cobot performance. Our non-parametric architecture outperforms the robot's default factory position controller in terms of accuracy.

A Lightweight convolutional neural network for nicotine prediction in tobacco by near-infrared spectroscopy.

The content of nicotine, a critical component of tobacco, significantly influences the quality of tobacco leaves. Near-infrared (NIR) spectroscopy is a widely used technique for rapid, non-destructive, and environmentally friendly analysis of nicotine levels in tobacco. In this paper, we propose a novel regression model, Lightweight one-dimensional convolutional neural network (1D-CNN), for predicting nicotine content in tobacco leaves using one-dimensional (1D) NIR spectral data and a deep learning approach with convolutional neural network (CNN). This study employed Savitzky-Golay (SG) smoothing to preprocess NIR spectra and randomly generate representative training and test datasets. Batch normalization was used in network regularization to reduce overfitting and improve the generalization performance of the Lightweight 1D-CNN model under a limited training dataset. The network structure of this CNN model consists of four convolutional layers to extract high-level features from the input data. The output of these layers is then fed into a fully connected layer, which uses a linear activation function to output the predicted numerical value of nicotine. After the comparison of the performance of multiple regression models, including support vector regression (SVR), partial least squares regression (PLSR), 1D-CNN, and Lightweight 1D-CNN, under the preprocessing method of SG smoothing, we found that the Lightweight 1D-CNN regression model with batch normalization achieved root mean square error (RMSE) of 0.14, coefficient of determination (R 2) of 0.95, and residual prediction deviation (RPD) of 5.09. These results demonstrate that the Lightweight 1D-CNN model is objective and robust and outperforms existing methods in terms of accuracy, which has the potential to significantly improve quality control processes in the tobacco industry by accurately and rapidly analyzing the nicotine content.

Crack identification using smart paint and machine learning

Information on the presence and location of cracks in civil structures can be precious to support operators in making decisions related to structural management and scheduling informed maintenance. This paper investigates the efficacy of supervised machine learning to solve the inverse electrical impedance tomography problem and to reconstruct the conductivity distribution of a piezoresistive sensing film. This film consists of a conductive paint applied onto structural components, and operators can use its conductivity distribution to identify crack sizes and locations in the underlying structure. A deep neural network is employed to reconstruct a dense conductivity distribution within the painted area by using only voltage measurements collected at sparse boundary locations. Since one of the most challenging aspects of using supervised learning tools for real-world applications is generating a representative training dataset, this paper presents a new approach to test the suitability of synthetic datasets built using a finite element model of the sensing film. Results are reported for four sensing specimens fabricated with two different techniques (i.e., using carbon nanotubes and graphene nanosheets, respectively). Crack-like damage is induced to the substrate of the sensing film and identified using the proposed machine learning technique. Promising results are obtained as compared to conventional methods.

Application of Model Data for Training the Classifier of Defects in Rail Bolt Holes in Ultrasonic Diagnostics

The task of searching for defects on defectograms of ultrasonic inspection of rails using machine learning methods is considered difficult due to the lack of a sufficient representative training data set. In this study, on the example of defects in bolt holes of rails, the possibility of synthesizing an effective classifier trained on artificial data obtained by mathematical modeling is shown. Experiments have shown a high accuracy of predicting samples of real (not model) data exceeding 99%.

Effective and efficient active learning for deep learning-based tissue image analysis.

Deep learning attained excellent results in digital pathology recently. A challenge with its use is that high quality, representative training datasets are required to build robust models. Data annotation in the domain is labor intensive and demands substantial time commitment from expert pathologists. Active learning (AL) is a strategy to minimize annotation. The goal is to select samples from the pool of unlabeled data for annotation that improves model accuracy. However, AL is a very compute demanding approach. The benefits for model learning may vary according to the strategy used, and it may be hard for a domain specialist to fine tune the solution without an integrated interface. We developed a framework that includes a friendly user interface along with run-time optimizations to reduce annotation and execution time in AL in digital pathology. Our solution implements several AL strategies along with our diversity-aware data acquisition (DADA) acquisition function, which enforces data diversity to improve the prediction performance of a model. In this work, we employed a model simplification strategy [Network Auto-Reduction (NAR)] that significantly improves AL execution time when coupled with DADA. NAR produces less compute demanding models, which replace the target models during the AL process to reduce processing demands. An evaluation with a tumor-infiltrating lymphocytes classification application shows that: (i) DADA attains superior performance compared to state-of-the-art AL strategies for different convolutional neural networks (CNNs), (ii) NAR improves the AL execution time by up to 4.3×, and (iii) target models trained with patches/data selected by the NAR reduced versions achieve similar or superior classification quality to using target CNNs for data selection. Source code: https://github.com/alsmeirelles/DADA.

Generalization of an Encoder-Decoder LSTM model for flood prediction in ungauged catchments

Flood prediction in ungauged catchments is usually conducted by hydrological models that are parameterized based on nearby and similar gauged catchments. As an alternative to this process-based modelling, deep learning (DL) models have demonstrated their ability for prediction in ungauged catchments (PUB) with high efficiency. Catchment characteristics, the number of gauged catchments, and their level of hydroclimatic heterogeneity in the training dataset used for model regionalization can directly affect the model’s performance. Here, we study the generalization ability of a DL model to these factors by applying an Encoder-Decoder Long Short-Term Memory neural network for a 6-hour lead-time runoff prediction in 35 mountainous catchments in China. By varying the available number of catchments and model settings with different training datasets, namely local, regional, and PUB models, we evaluated the generalization ability of our model. We found that both quantity (i.e. number of gauged catchments available) and heterogeneity of the training dataset used for the DL model are important for improving model performance in the PUB context, due to a data synergy effect. The assessment of the sensitivity to catchment characteristics showed that the model performance is mainly correlated to the local hydro-climatic conditions; the more arid the region, the more likely it is to have a poor model performance for prediction in ungauged catchments. The results suggest that the regional ED-LSTM model is a promising method to predict streamflow from rainfall inputs in PUB, and outline the need for preparing a representative training dataset.

Efficient data acquisition and training of collisional-radiative model artificial neural network surrogates through adaptive parameter space sampling

Effective plasma transport modeling of magnetically confined fusion devices relies on having an accurate understanding of the ion composition and radiative power losses of the plasma. Generally, these quantities can be obtained from solutions of a collisional-radiative (CR) model at each time step within a plasma transport simulation. However, even compact, approximate CR models can be computationally onerous to evaluate, and in-situ evaluation of these models within a larger plasma transport code can lead to a rigid bottleneck. As a way to bypass this bottleneck, we propose deploying artificial neural network (ANN) surrogates to allow rapid evaluation of the necessary plasma quantities. However, one issue with training an accurate ANN surrogate is the reliance on a sufficiently large and representative training and validation data set, which can be time-consuming to generate. In this work we explore a data-driven active learning and training routine to allow autonomous adaptive sampling of the problem parameter space to ensure a sufficiently large and meaningful set of training data is assembled for the network training. As a result, we can demonstrate approximately order-of-magnitude savings in required training data samples to produce an accurate surrogate.

Adaptive Feedback Convolutional‐Neural‐Network‐Based High‐Resolution Reflection‐Waveform Inversion

AbstractFull‐waveform inversion (FWI) applies non‐linear optimization to estimate the velocity model by fitting the observed seismic data. With a smooth starting velocity model, FWI mainly inverts for the shallower background velocity model by fitting the observed direct, diving and refracted data, and updates the interfaces by fitting the observed reflected data. As the deeper parts of background velocity model cannot be effectively updated by fitting the reflected data in FWI, the deeper interfaces are less accurate than the shallower interfaces. To update the deeper background velocity model, many reflection‐waveform inversion (RWI) algorithms were proposed to separate the tomographic and migration components from the reflection‐related gradient. We propose a convolutional‐neural‐network‐based reflection‐waveform inversion (CNN‐RWI) to repeatedly apply the iteratively updated convolutional neural network (CNN) to predict the true velocity model from the smooth starting velocity model (the tomographic components), and the high‐resolution migration image (the migration components). The CNN is iteratively updated by the more representative training data set, which is obtained from the latest CNN‐predicted velocity model by the proposed spatially constrained divisive hierarchical k‐means parcellation method. The more representative the training velocity models are, the more accurate the CNN‐predicted velocity model becomes. Synthetic examples using different portions of the Marmousi2 P‐wave velocity model show that CNN‐RWI inverts for both the shallower and deeper parts of velocity models more accurately than the conjugate‐gradient FWI (CG‐FWI). Both the CNN‐RWI and the CG‐FWI are sensitive to the accuracy of the starting velocity model and the complexity of the unknown true velocity model.

Bi-Long Short-Term Memory Networks for Radio Frequency Based Arrival Time Detection of Partial Discharge Signals

Partial discharge (PD) monitoring of electrical substations could provide early warning of insulation failures. Among the various technologies, Radio Frequency (RF) based PD monitoring system could be a promising solution. The RF-based monitoring system detects PD sources in the substation and can also localise the PD sources. The time difference of arrival (TDOA) based PD localisation system primarily require arrival time (AT) of the impulsive RF signal. Though many localisation algorithms have been proposed in the recent past to overcome the TDOA estimation errors, less attention has been given to the accurate estimation of RF PD signal arrival time. This paper presents the AT's automatic labelling in the RF PD signal using Bi-Long Short-Term Memory (Bi-LSTM) network applied on the continuous wavelet transformed (CWT) signal. Further, it also shows PD signal augmentation to overcome the problem of limited representative training data set. The behavior of the radiated RF signals is influenced by many factors and has almost stochastic characteristics. The proposed system has been validated with laboratory-based experimental signals and the data set obtained from different electrical substations. The results show that the improved performance is obtained from the combination of a multilayer Bi-LSTM model and an augmented training set.

A Deep-Learning intelligent system incorporating data augmentation for Short-Term voltage stability assessment of power systems

Facing the difficulty of expensive and trivial data collection and annotation, how to make a deep learning-based short-term voltage stability assessment (STVSA) model work well on a small training dataset is a challenging and urgent problem. Although a big enough dataset can be directly generated by contingency simulation, this data generation process is usually cumbersome and inefficient; while data augmentation provides a low-cost and efficient way to artificially inflate the representative and diversified training datasets with label preserving transformations. In this respect, this paper proposes a novel deep-learning intelligent system incorporating data augmentation for STVSA of power systems. First, due to the unavailability of reliable quantitative criteria to judge the stability status for a specific power system, semi-supervised cluster learning is leveraged to obtain labeled samples in an original small dataset. Second, to make deep learning applicable to the small dataset, conditional least squares generative adversarial networks (LSGAN)-based data augmentation is introduced to expand the original dataset via artificially creating additional valid samples. Third, to extract temporal dependencies from the post-disturbance dynamic trajectories of a system, a bi-directional gated recurrent unit with attention mechanism based assessment model is established, which bi-directionally learns the significant time dependencies and automatically allocates attention weights. The test results demonstrate the presented approach manages to achieve better accuracy and a faster response time with original small datasets. Besides classification accuracy, this work employs statistical measures to comprehensively examine the performance of the proposal.

The Use of Synthetic IMU Signals in the Training of Deep Learning Models Significantly Improves the Accuracy of Joint Kinematic Predictions.

Gait analysis based on inertial sensors has become an effective method of quantifying movement mechanics, such as joint kinematics and kinetics. Machine learning techniques are used to reliably predict joint mechanics directly from streams of IMU signals for various activities. These data-driven models require comprehensive and representative training datasets to be generalizable across the movement variability seen in the population at large. Bottlenecks in model development frequently occur due to the lack of sufficient training data and the significant time and resources necessary to acquire these datasets. Reliable methods to generate synthetic biomechanical training data could streamline model development and potentially improve model performance. In this study, we developed a methodology to generate synthetic kinematics and the associated predicted IMU signals using open source musculoskeletal modeling software. These synthetic data were used to train neural networks to predict three degree-of-freedom joint rotations at the hip and knee during gait either in lieu of or along with previously measured experimental gait data. The accuracy of the models’ kinematic predictions was assessed using experimentally measured IMU signals and gait kinematics. Models trained using the synthetic data out-performed models using only the experimental data in five of the six rotational degrees of freedom at the hip and knee. On average, root mean square errors in joint angle predictions were improved by 38% at the hip (synthetic data RMSE: 2.3°, measured data RMSE: 4.5°) and 11% at the knee (synthetic data RMSE: 2.9°, measured data RMSE: 3.3°), when models trained solely on synthetic data were compared to measured data. When models were trained on both measured and synthetic data, root mean square errors were reduced by 54% at the hip (measured + synthetic data RMSE: 1.9°) and 45% at the knee (measured + synthetic data RMSE: 1.7°), compared to measured data alone. These findings enable future model development for different activities of clinical significance without the burden of generating large quantities of gait lab data for model training, streamlining model development, and ultimately improving model performance.

Interpreting Deep Learning Studies in Glaucoma: Unresolved Challenges.

Deep learning algorithms as tools for automated image classification have recently experienced rapid growth in imaging-dependent medical specialties, including ophthalmology. However, only a few algorithms tailored to specific health conditions have been able to achieve regulatory approval for autonomous diagnosis. There is now an international effort to establish optimized thresholds for algorithm performance benchmarking in a rapidly evolving artificial intelligence field. This review examines the largest deep learning studies in glaucoma, with special focus on identifying recurrent challenges and limitations within these studies which preclude widespread clinical deployment. We focus on the 3 most common input modalities when diagnosing glaucoma, namely, fundus photographs, spectral domain optical coherence tomography scans, and standard automated perimetry data. We then analyze 3 major challenges present in all studies: defining the algorithm output of glaucoma, determining reliable ground truth datasets, and compiling representative training datasets.

Machine Learning of Coarse-Grained Models for Organic Molecules and Polymers: Progress, Opportunities, and Challenges.

Machine learning (ML) has emerged as one of the most powerful tools transforming all areas of science and engineering. The nature of molecular dynamics (MD) simulations, complex and time-consuming calculations, makes them particularly suitable for ML research. This review article focuses on recent advancements in developing efficient and accurate coarse-grained (CG) models using various ML methods, in terms of regulating the coarse-graining process, constructing adequate descriptors/features, generating representative training data sets, and optimization of the loss function. Two classes of the CG models are introduced: bottom-up and top-down CG methods. To illustrate these methods and demonstrate the open methodological questions, we survey several important principles in constructing CG models and how these are incorporated into ML methods and improved with specific learning techniques. Finally, we discuss some key aspects of developing machine-learned CG models with high accuracy and efficiency. Besides, we describe how these aspects are tackled in state-of-the-art methods and which remain to be addressed in the near future. We expect that these machine-learned CG models can address thermodynamic consistent, transferable, and representative issues in classical CG models.