Entire Training Dataset Research Articles

Training-free approach to constructing ensemble of local experts

Conventional machine learning methods assume that a central server holds the entire training dataset for training a global model. However, in many real-world situations, datasets are collected from individual local nodes in distributed environments. Gathering these datasets often poses challenges owing to various practical constraints such as data security and privacy issues, storage limits, and transmission costs. When data distributions differ among local nodes, local experts trained on individual local nodes exhibit poor generalization performance for new instances. In this study, we propose a training-free method for constructing an ensemble of local experts without accessing datasets. For a query instance, we apply uncertainty quantification and out-of-distribution (OOD) detection to calculate the uncertainty and OOD scores of the predictions from individual models. We then determine the weights of the models by assigning a lower weight to those with higher uncertainty and OOD scores. Using these weights, the predictions of the individual models are aggregated to obtain a final prediction. In contrast to existing methods that impose specific application requirements, the proposed method can be applied in situations in which only individual predictions from local models are available for use when making a prediction for a query instance. We evaluate the effectiveness of the proposed method using image classification benchmark datasets.

Optimizing motor imagery BCI models with hard trials removal and model refinement

Deep learning models have demonstrated remarkable performance in the classification of motor imagery BCI systems. However, these models exhibit sensitivity to challenging trials, often called hard trials, leading to performance degradation. In this paper, we address this issue by proposing two novel methods for identifying and mitigating the impact of hard trials on model performance. The first method leverages model prediction scores to discern hard trials. The second approach employs a quantitative explainable artificial intelligence (XAI) approach, enabling a more transparent and interpretable means of hard trial identification. The identified hard trials are removed from the entire motor imagery training and validation dataset, and the deep learning model is further re-trained using the dataset without hard trials. To evaluate the efficacy of these proposed methods, experiments were conducted on the Open BMI dataset. The results for hold-out analysis show that the proposed quantitative XAI- based hard trial removal method has statistically improved the average classification accuracy of the baseline deep CNN model from 63.77% to 68.70%, with p-value = 7.66e−11 for the subject-specific MI classification. Additionally, analyzing the scalp map representing the average relevance scores of correctly classified trials compared to a misclassified trial provides a deeper insight into identifying hard trials. The results indicate that the proposed quantitaive-based XAI approach outperformes the prediction-score-based approach in hard trial identification.

Application of a U-Net Neural Network to the Puccinia sorghi-Maize Pathosystem.

Computer vision approaches to analyze plant disease data can be both faster and more reliable than traditional, manual methods. However, the requirement of manually annotating training data for the majority of machine learning applications can present a challenge for pipeline development. Here, we describe a machine learning approach to quantify Puccinia sorghi incidence on maize leaves utilizing U-Net convolutional neural network models. We analyzed several U-Net models with increasing amounts of training image data, either randomly chosen from a large data pool or randomly chosen from a subset of disease time course data. As the training dataset size increases, the models perform better, but the rate of performance decreases. Additionally, the use of a diverse training dataset can improve model performance and reduce the amount of annotated training data required for satisfactory performance. Models with as few as 48 whole-leaf training images are able to replicate the ground truth results within our testing dataset. The final model utilizing our entire training dataset performs similarly to our ground truth data, with an intersection over union value of 0.5002 and an F1 score of 0.6669. This work illustrates the capacity of U-Nets to accurately answer real-world plant pathology questions related to quantification and estimation of plant disease symptoms. [Formula: see text] Copyright © 2024 The Author(s). This is an open access article distributed under the CC BY-NC-ND 4.0 International license.

Compact Modeling of Advanced Gate-All-Around Nanosheet FETs Using Artificial Neural Network.

As the architecture of logic devices is evolving towards gate-all-around (GAA) structure, research efforts on advanced transistors are increasingly desired. In order to rapidly perform accurate compact modeling for these ultra-scaled transistors with the capability to cover dimensional variations, neural networks are considered. In this paper, a compact model generation methodology based on artificial neural network (ANN) is developed for GAA nanosheet FETs (NSFETs) at advanced technology nodes. The DC and AC characteristics of GAA NSFETs with various physical gate lengths (Lg), nanosheet widths (Wsh) and thicknesses (Tsh), as well as different gate voltages (Vgs) and drain voltages (Vds) are obtained through TCAD simulations. Subsequently, a high-precision ANN model architecture is evaluated. A systematical study on the impacts of ANN size, activation function, learning rate, and epoch (the times of complete pass through the entire training dataset) on the accuracy of ANN models is conducted, and a shallow neural network configuration for generating optimal ANN models is proposed. The results clearly show that the optimized ANN model can reproduce the DC and AC characteristics of NSFETs very accurately with a fitting error (MSE) of 0.01.

Two-Step Hyperparameter Optimization Method: Accelerating Hyperparameter Search by Using a Fraction of a Training Dataset

Abstract Hyperparameter optimization (HPO) is an important step in machine learning (ML) model development, but common practices are archaic—primarily relying on manual or grid searches. This is partly because adopting advanced HPO algorithms introduces added complexity to the workflow, leading to longer computation times. This poses a notable challenge to ML applications, as suboptimal hyperparameter selections curtail the potential of ML model performance, ultimately obstructing the full exploitation of ML techniques. In this article, we present a two-step HPO method as a strategic solution to curbing computational demands and wait times, gleaned from practical experiences in applied ML parameterization work. The initial phase involves a preliminary evaluation of hyperparameters on a small subset of the training dataset, followed by a reevaluation of the top-performing candidate models postretraining with the entire training dataset. This two-step HPO method is universally applicable across HPO search algorithms, and we argue it has attractive efficiency gains. As a case study, we present our recent application of the two-step HPO method to the development of neural network emulators for aerosol activation. Although our primary use case is a data-rich limit with many millions of samples, we also find that using up to 0.0025% of the data—a few thousand samples—in the initial step is sufficient to find optimal hyperparameter configurations from much more extensive sampling, achieving up to 135× speedup. The benefits of this method materialize through an assessment of hyperparameters and model performance, revealing the minimal model complexity required to achieve the best performance. The assortment of top-performing models harvested from the HPO process allows us to choose a high-performing model with a low inference cost for efficient use in global climate models (GCMs).

Combining Fuzzy Partitioning and Incremental Methods to Construct a Scalable Decision Tree on Large Datasets

The Decision tree algorithm is a very popular classifier for reasoning through recursive partitioning of the data space. To choose the best attributes for splitting, the range of each continuous attribute should be split into two or more intervals. Then partitioning criteria are calculated for each value. Fuzzy partitioning can be used to reduce sensitivity to noise and increase tree stability. Also, tree-building algorithms face memory limitations as they need to keep the entire training dataset in the main memory. In this paper, we introduced a fuzzy decision tree approach based on fuzzy sets. To avoid storing the entire training dataset in the main memory and overcome the memory limitations, the algorithm incrementally builds FDTs. Membership functions are automatically generated. The Fuzzy Information Gain (FIG) is then used as the fast split attribute selection criterion, and leaf expansion is performed only on the instances stored in it. The efficiency of this algorithm is examined in terms of accuracy and tree complexity. The results show that the proposed algorithm can overcome memory limitations and balance accuracy and complexity while reducing the complexity of the tree.

Deep neural network based tissue deconvolution of circulating tumor cell RNA

Prior research has shown that the deconvolution of cell-free RNA can uncover the tissue origin. The conventional deconvolution approaches rely on constructing a reference tissue-specific gene panel, which cannot capture the inherent variation present in actual data. To address this, we have developed a novel method that utilizes a neural network framework to leverage the entire training dataset. Our approach involved training a model that incorporated 15 distinct tissue types. Through one semi-independent and two complete independent validations, including deconvolution using a semi in silico dataset, deconvolution with a custom normal tissue mixture RNA-seq data, and deconvolution of longitudinal circulating tumor cell RNA-seq (ctcRNA) data from a cancer patient with metastatic tumors, we demonstrate the efficacy and advantages of the deep-learning approach which were exerted by effectively capturing the inherent variability present in the dataset, thus leading to enhanced accuracy. Sensitivity analyses reveal that neural network models are less susceptible to the presence of missing data, making them more suitable for real-world applications. Moreover, by leveraging the concept of organotropism, we applied our approach to trace the migration of circulating tumor cell-derived RNA (ctcRNA) in a cancer patient with metastatic tumors, thereby highlighting the potential clinical significance of early detection of cancer metastasis.

Analysis of alternative strategies applied to Naïve-Bayes classifier into the recognition of electrofacies: Application in well-log data at Recôncavo Basin, North-East Brazil

This paper is concerned with the applicability of different strategies to improve the definition of prior probabilities and/or likelihoods of naïve Bayes (NB) classifiers. Standard NB method computes likelihoods and priors by means of normal distributions and evaluation of the entire training data-set, respectively. NB is one of the most prolific classification methods in data mining and machine learning. Despite decent efficiency facing good training data, NB classifiers present the intriguing assumption of conditional independence between the attributes. Several algorithms have been proposed to improve the effectiveness of NB classifiers by inserting discriminant approaches into its generative structure. From a reliability perspective, the standard NB approach might not explore the real capabilities of NB classifiers facing the lithologic classification problem. To cover such distrust, a novel approach considering four particular strategies are suggested and compared to the standard NB classification outcomes. At first, a kernel density estimation (KDE) is considered to ameliorate the likelihood models. We also apply the NB classifier in parts by separating the training data-set in individual wells in a committee architecture framework. A tuning strategy is also considered for automatic estimation of prior probabilities in an optimization-scheme. Another novel alternative, named as CRC, adapted to the standard NB classifier consists in defining priors based on depth zones from regional stratigraphic information in which to apply NB classifier. We prepare an extensive statistical investigation, based on precision, recall, classification errors, fscores and confusion matrices to bespeak the most relevant NB strategy for classification of electrofacies. Despite the decent classification outcomes for all above-mentioned strategies, CRC can be considered, by a narrow margin, the most prolific method to be applied as an improvement of the standard NB to classify rock units. Tests are performed on a validation well (i.e., 7-MP-50D-BA) of Massapê Field, in Recôncavo Basin, northeast Brazil to highlight the classification particularities provided by the improved strategy. A significant improvement in the classification of sandstones (i.e., from 68 % to 83 % accuracy) is observed, according to the confusion matrix analysis. Additionally, a minimal decreasing is observed into the classification of shales and slurries (i.e., from 92% to 90% and to 63% to 56%, respectively), which is acceptable according to the fscores and errors. This aspect reinforces the relevance of using the NB classifier jointly with previous geologic information to optimize the lithologic classification.

Deep-Learning-Based Whole-Lung and Lung-Lesion Quantification Despite Inconsistent Ground Truth: Application to Computerized Tomography in SARS-CoV-2 Nonhuman Primate Models

Animal modeling of infectious diseases such as coronavirus disease 2019 (COVID-19) is important for exploration of natural history, understanding of pathogenesis, and evaluation of countermeasures. Preclinical studies enable rigorous control of experimental conditions as well as pre-exposure baseline and longitudinal measurements, including medical imaging, that are often unavailable in the clinical research setting. Computerized tomography (CT) imaging provides important diagnostic, prognostic, and disease characterization to clinicians and clinical researchers. In that context, automated deep-learning systems for the analysis of CT imaging have been broadly proposed, but their practical utility has been limited. Manual outlining of the ground truth (i.e., lung-lesions) requires accurate distinctions between abnormal and normal tissues that often have vague boundaries and is subject to reader heterogeneity in interpretation. Indeed, this subjectivity is demonstrated as wide inconsistency in manual outlines among experts and from the same expert. The application of deep-learning data-science tools has been less well-evaluated in the preclinical setting, including in nonhuman primate (NHP) models of severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infection/COVID-19, in which the translation of human-derived deep-learning tools is challenging. The automated segmentation of the whole lung and lung lesions provides a potentially standardized and automated method to detect and quantify disease. We used deep-learning-based quantification of the whole lung and lung lesions on CT scans of NHPs exposed to SARS-CoV-2. We proposed a novel multi-model ensemble technique to address the inconsistency in the ground truths for deep-learning-based automated segmentation of the whole lung and lung lesions. Multiple models were obtained by training the convolutional neural network (CNN) on different subsets of the training data instead of having a single model using the entire training dataset. Moreover, we employed a feature pyramid network (FPN), a CNN that provides predictions at different resolution levels, enabling the network to predict objects with wide size variations. We achieved an average of 99.4 and 60.2% Dice coefficients for whole-lung and lung-lesion segmentation, respectively. The proposed multi-model FPN outperformed well-accepted methods U-Net (50.5%), V-Net (54.5%), and Inception (53.4%) for the challenging lesion-segmentation task. We show the application of segmentation outputs for longitudinal quantification of lung disease in SARS-CoV-2-exposed and mock-exposed NHPs. Deep-learning methods should be optimally characterized for and targeted specifically to preclinical research needs in terms of impact, automation, and dynamic quantification independently from purely clinical applications.

Structure Destruction and Content Combination for Generalizable Anti-Spoofing

In securing the face verification systems, prior face anti-spoofing studies excavate hidden cues in original images to discriminate real persons and diverse attacks with the assistance of auxiliary supervision. However, limited by several inherent shortcomings in their training process: 1) Neglect of the multi-scale nature of spoof cues; 2) Complete integral structure in a single image; 3) Implicit subdomains in the whole dataset, these methods are weak to mine comprehensive spoof patterns and may stick on memorization of the entire training dataset, incurring overfitting. In this paper, we propose a new framework named Destruction and Combination Network (DCN) including Multi-scale Representation Extraction Module, Structure Destruction Module, and Content Combination Module to address these limitations respectively. The first mechanism exploits multi-scale representation to learn spoof cues comprehensively. The second one destroys images into patches to construct non-structural inputs, and the last scheme recombines patches from different subdomains or classes into a mixup construction. Based on the above splitting-and-splicing operation, we further introduce Local Relation Modeling Module to model the second-order relationship between patches. To show the generalizable capacity of the proposed framework, besides simple intra-dataset testing, we test our method in cross-domain, cross-content, and cross-attack scenarios. Extensive experiments on different scenarios demonstrate the reliability of our method against state-of-the-art competitors.

A Survey of the Recent Trends in Deep Learning Based Malware Detection

Monitoring Indicators of Compromise (IOC) leads to malware detection for identifying malicious activity. Malicious activities potentially lead to a system breach or data compromise. Various tools and anti-malware products exist for the detection of malware and cyberattacks utilizing IOCs, but all have several shortcomings. For instance, anti-malware systems make use of malware signatures, requiring a database containing such signatures to be constantly updated. Additionally, this technique does not work for zero-day attacks or variants of existing malware. In the quest to fight zero-day attacks, the research paradigm shifted from primitive methods to classical machine learning-based methods. Primitive methods are limited in catering to anti-analysis techniques against zero-day attacks. Hence, the direction of research moved towards methods utilizing classic machine learning, however, machine learning methods also come with certain limitations. They may include but not limited to the latency/lag introduced by feature-engineering phase on the entire training dataset as opposed to the real-time analysis requirement. Likewise, additional layers of data engineering to cater to the increasing volume of data introduces further delays. It led to the use of deep learning-based methods for malware detection. With the speedy occurrence of zero-day malware, researchers chose to experiment with few shot learning so that reliable solutions can be produced for malware detection with even a small amount of data at hand for training. In this paper, we surveyed several possible strategies to support the real-time detection of malware and propose a hierarchical model to discover security events or threats in real-time. A key focus in this survey is on the use of Deep Learning-based methods. Deep Learning based methods dominate this research area by providing automatic feature engineering, the capability of dealing with large datasets, enabling the mining of features from limited data samples, and supporting one-shot learning. We compare Deep Learning-based approaches with conventional machine learning based approaches and primitive (statistical analysis based) methods commonly reported in the literature.

Pseudo loss active learning for deep visual tracking

In visual tracking tasks, the training data are commonly composed of a large number of video sequences and each frame in the sequences needs to be labeled manually, which is labor-intensive and time-consuming. In addition, considering the similarity among the consecutive frames in the same sequence, there is significant redundancy in the training data. To address these problems, a novel pseudo loss active learning (PLAL) method is developed in this paper. PLAL aims to select the most informative and least redundant data for training to reduce the cost of labeling and maintain competitive tracking results simultaneously. Firstly, the Gaussian distribution based pseudo label is generated for the unlabeled candidates based on the tracking model which is initially trained on a small amount of training data. Then, the pseudo loss based on cross entropy is designed to compute the difference between the pseudo label and the target response map. The pseudo loss measures the uncertainty of the target spatial context which is used as the informativeness criterion of the image frame for selection. Meanwhile, a sampling interval threshold and a temporal penalty are employed for frame selection to avoid drastic variation in target appearance and reduce the redundancy within the consecutive candidate frames. Only the selected frames are labeled by the oracle (human expert) and then added to the training data. Extensive experiments on public benchmarks (OTB2013, OTB2015, VOT2018, UAV123, GOT-10K, TrackingNet, LaSOT, OxUvA and TLP) demonstrate that PLAL method outperforms the baseline and other recent active learning approaches. With only 3% of labeled data from the training dataset, PLAL reaches competitive performance (98-100%) compared to the model trained on the entire training dataset.

A comparison of approaches to improve worst-case predictive model performance over patient subpopulations

Predictive models for clinical outcomes that are accurate on average in a patient population may underperform drastically for some subpopulations, potentially introducing or reinforcing inequities in care access and quality. Model training approaches that aim to maximize worst-case model performance across subpopulations, such as distributionally robust optimization (DRO), attempt to address this problem without introducing additional harms. We conduct a large-scale empirical study of DRO and several variations of standard learning procedures to identify approaches for model development and selection that consistently improve disaggregated and worst-case performance over subpopulations compared to standard approaches for learning predictive models from electronic health records data. In the course of our evaluation, we introduce an extension to DRO approaches that allows for specification of the metric used to assess worst-case performance. We conduct the analysis for models that predict in-hospital mortality, prolonged length of stay, and 30-day readmission for inpatient admissions, and predict in-hospital mortality using intensive care data. We find that, with relatively few exceptions, no approach performs better, for each patient subpopulation examined, than standard learning procedures using the entire training dataset. These results imply that when it is of interest to improve model performance for patient subpopulations beyond what can be achieved with standard practices, it may be necessary to do so via data collection techniques that increase the effective sample size or reduce the level of noise in the prediction problem.

Fabric defect classification using prototypical network of few-shot learning algorithm

Computer-vision technology plays a vital role in automated fabric defect classification. In this paper, a novel prototypical network is presented for improving the fabric defect classification performance, especially in the case of an imbalanced distribution over the number of class samples. A traditional neural network (such as a convolutional neural network) usually inputs a batch of samples each time in training until the entire training dataset is covered, and thus it is not robust to cope with imbalanced data. The proposed network follows an N-way K-shot paradigm to split the training set into a support set and query set, and thereby forces the number of samples within each class to be uniformly distributed. The support set is used to learn the common knowledge of each class, whereas the query set is utilized to fine-tune the model parameters gained from the respective support set. The prototype of each class in the support set is computed as the representation of the class. For samples in the query set, the loss function is designed to match them with the corresponding prototypes as accurately as possible. In addition, the class activation mapping is used to visualize and interpret the discriminative regions of interest most relevant to specific defect classes. The classification performance of the proposed method is tested against five existing models on a labeled dataset of fabric defect images collected by a commercial inspection system. The proposed method achieves the highest classification accuracy (96.04%) over seven defect categories among the six tested methods.

A novel pure data-selection framework for day-ahead wind power forecasting

Numerical weather prediction (NWP) data possess internal inaccuracies, such as low NWP wind speed corresponding to high actual wind power generation. This study is intended to reduce the negative effects of such inaccuracies by proposing a pure data-selection framework (PDF) to choose useful data prior to modeling, thus improving the accuracy of day-ahead wind power forecasting. Briefly, we convert an entire NWP training dataset into many small subsets and then select the best subset combination via a validation set to build a forecasting model. Although a small subset can increase selection flexibility, it can also produce billions of subset combinations, resulting in computational issues. To address this problem, we incorporated metamodeling and optimization steps into PDF. We then proposed a design and analysis of the computer experiments-based metamodeling algorithm and heuristic-exhaustive search optimization algorithm, respectively.Experimental results demonstrate that (1) it is necessary to select data before constructing a forecasting model; (2) using a smaller subset will likely increase selection flexibility, leading to a more accurate forecasting model; (3) PDF can generate a better training dataset than similarity-based data selection methods (e.g., K-means and support vector classification); and (4) choosing data before building a forecasting model produces a more accurate forecasting model compared with using a machine learning method to construct a model directly.

Impact of Intraoperative Data on Risk Prediction for Mortality After Intra-Abdominal Surgery.

Risk prediction models for postoperative mortality after intra-abdominal surgery have typically been developed using preoperative variables. It is unclear if intraoperative data add significant value to these risk prediction models. With IRB approval, an institutional retrospective cohort of intra-abdominal surgery patients in the 2005 to 2015 American College of Surgeons National Surgical Quality Improvement Program was identified. Intraoperative data were obtained from the electronic health record. The primary outcome was 30-day mortality. We evaluated the performance of machine learning algorithms to predict 30-day mortality using: 1) baseline variables and 2) baseline + intraoperative variables. Algorithms evaluated were: 1) logistic regression with elastic net selection, 2) random forest (RF), 3) gradient boosting machine (GBM), 4) support vector machine (SVM), and 5) convolutional neural networks (CNNs). Model performance was evaluated using the area under the receiver operator characteristic curve (AUROC). The sample was randomly divided into a training/testing split with 80%/20% probabilities. Repeated 10-fold cross-validation identified the optimal model hyperparameters in the training dataset for each model, which were then applied to the entire training dataset to train the model. Trained models were applied to the test cohort to evaluate model performance. Statistical significance was evaluated using P < .05. The training and testing cohorts contained 4322 and 1079 patients, respectively, with 62 (1.4%) and 15 (1.4%) experiencing 30-day mortality, respectively. When using only baseline variables to predict mortality, all algorithms except SVM (area under the receiver operator characteristic curve [AUROC], 0.83 [95% confidence interval {CI}, 0.69-0.97]) had AUROC >0.9: GBM (AUROC, 0.96 [0.94-1.0]), RF (AUROC, 0.96 [0.92-1.0]), CNN (AUROC, 0.96 [0.92-0.99]), and logistic regression (AUROC, 0.95 [0.91-0.99]). AUROC significantly increased with intraoperative variables with CNN (AUROC, 0.97 [0.96-0.99]; P = .047 versus baseline), but there was no improvement with GBM (AUROC, 0.97 [0.95-0.99]; P = .3 versus baseline), RF (AUROC, 0.96 [0.93-1.0]; P = .5 versus baseline), and logistic regression (AUROC, 0.94 [0.90-0.99]; P = .6 versus baseline). Postoperative mortality is predicted with excellent discrimination in intra-abdominal surgery patients using only preoperative variables in various machine learning algorithms. The addition of intraoperative data to preoperative data also resulted in models with excellent discrimination, but model performance did not improve.

Exploratory Radiomic Analysis of Conventional vs. Quantitative Brain MRI: Toward Automatic Diagnosis of Early Multiple Sclerosis.

Conventional magnetic resonance imaging (cMRI) is poorly sensitive to pathological changes related to multiple sclerosis (MS) in normal-appearing white matter (NAWM) and gray matter (GM), with the added difficulty of not being very reproducible. Quantitative MRI (qMRI), on the other hand, attempts to represent the physical properties of tissues, making it an ideal candidate for quantitative medical image analysis or radiomics. We therefore hypothesized that qMRI-based radiomic features have added diagnostic value in MS compared to cMRI. This study investigated the ability of cMRI (T1w) and qMRI features extracted from white matter (WM), NAWM, and GM to distinguish between MS patients (MSP) and healthy control subjects (HCS). We developed exploratory radiomic classification models on a dataset comprising 36 MSP and 36 HCS recruited in CHU Liege, Belgium, acquired with cMRI and qMRI. For each image type and region of interest, qMRI radiomic models for MS diagnosis were developed on a training subset and validated on a testing subset. Radiomic models based on cMRI were developed on the entire training dataset and externally validated on open-source datasets with 167 HCS and 10 MSP. Ranked by region of interest, the best diagnostic performance was achieved in the whole WM. Here the model based on magnetization transfer imaging (a type of qMRI) features yielded a median area under the receiver operating characteristic curve (AUC) of 1.00 in the testing sub-cohort. Ranked by image type, the best performance was achieved by the magnetization transfer models, with median AUCs of 0.79 (0.69–0.90, 90% CI) in NAWM and 0.81 (0.71–0.90) in GM. The external validation of the T1w models yielded an AUC of 0.78 (0.47–1.00) in the whole WM, demonstrating a large 95% CI and a low sensitivity of 0.30 (0.10–0.70). This exploratory study indicates that qMRI radiomics could provide efficient diagnostic information using NAWM and GM analysis in MSP. T1w radiomics could be useful for a fast and automated check of conventional MRI for WM abnormalities once acquisition and reconstruction heterogeneities have been overcome. Further prospective validation is needed, involving more data for better interpretation and generalization of the results.

Modeling of turbulent flames with the large eddy simulation–probability density function (LES–PDF) approach, stochastic fields, and artificial neural networks

This work proposes a chemical mechanism tabulation method using artificial neural networks (ANNs) for turbulent combustion simulations. The method is employed here in the context of the Large-Eddy Simulation (LES)–Probability Density Function (PDF) approach and the method of stochastic fields for numerical solution, but can also be employed in other methods featuring real-time integration of chemical kinetics. The focus of the paper is on exploring an ANN architecture aiming at improved generalization, which uses a single multilayer perceptron (MLP) for each species over the entire training dataset. This method is shown to outperform previous approaches which take advantage of specialization by clustering the composition space using the Self-Organizing Map (SOM). The ANN training data are generated using the canonical combustion problem of igniting/extinguishing one-dimensional laminar flamelets with a detailed methane combustion mechanism, before being augmented with randomly generated data to produce a hybrid random/flamelet dataset with improved composition space coverage. The ANNs generated in this study are applied to the LES of a turbulent non-premixed CH4/air flame, Sydney flame L. The transported PDF approach is used for reaction source term closure, while numerical solution is obtained using the method of stochastic fields. Very good agreement is observed between direct integration (DI) and the ANNs, meaning that the ANNs can successfully replace the integration of chemical kinetics. The time taken for the reaction source computation is reduced 18-fold, which means that LES–PDF simulations with comprehensive mechanisms can be performed on modest computing resources.

PA-Cache: Evolving Learning-Based Popularity- Aware Content Caching in Edge Networks

As ubiquitous and personalized services are growing boomingly, an increasingly large amount of traffic is generated over the network by massive mobile devices. As a result, content caching is gradually extending to network edges to provide low-latency services, improve quality of service, and reduce redundant data traffic. Compared to the conventional content delivery networks, caches in edge networks with smaller sizes usually have to accommodate more bursty requests. In this article, we propose an evolving learning-based content caching policy, named PA-Cache in edge networks. It adaptively learns time-varying content popularity and determines which contents should be replaced when the cache is full. Unlike conventional deep neural networks (DNNs), which learn a fine-tuned but possibly outdated or biased prediction model using the entire training dataset with high computational complexity, PA-Cache weighs a large set of content features and trains the multi-layer recurrent neural network from shallow to deeper when more requests arrive over time. We extensively evaluate the performance of our proposed PA-Cache on real-world traces from a large online video-on-demand service provider. The results show that PA-Cache outperforms existing popular caching algorithms and approximates the optimal algorithm with only a 3.8% performance gap when the cache percentage is 1.0%. PA-Cache also significantly reduces the computational cost compared to conventional DNN-based approaches.

A Few-Shot Learning Method Using Feature Reparameterization and Dual-Distance Metric Learning for Object Re-Identification

Many object re-identification (Re-ID) methods that depend on large-scale training datasets have been proposed in recent years. However, the performance of these methods degrades dramatically when insufficient training data are available. To address this challenging problem, we propose a few-shot object re-identification (FSOR) method that enhances the generalization and discrimination abilities of object Re-ID models trained on small datasets. This method applies two novel techniques: reparameterization for feature vectors and dual-distance metric learning. The reparameterization mechanism transforms the primary feature vector of each input image into a Gaussian distribution to enhance the robustness of the FSOR method when performing object Re-ID tasks. The dual-distance metric learning technique, called H&C learning, considers both the hard mining distance and the center-point distance between each query sample and each support set of different object identities. H&C learning extracts the characteristics of the entire training dataset more precisely than other approaches and thus improves the discriminative abilities of object Re-ID models. Extensive experiments on both person and vehicle Re-ID datasets, such as Market-1501, DukeMTMC-ReID, CUHK03, and VeRi-776, show that the FSOR method has improved performance and outperforms state-of-the-art methods when the amount of labeled training data is small.