A Supervised Learning Method for Improving the Generalization of Speaker Verification Systems by Learning Metrics from a Mean Teacher

The majority of recent speaker verification tasks are studied under open-set evaluation scenarios considering real-world conditions. The characteristics of these tasks imply that the generalization towards unseen speakers is a critical capability. Thus, this study aims to improve the generalization of the system for the performance enhancement of speaker verification. To achieve this goal, we propose a novel supervised-learning-method-based speaker verification system using the mean teacher framework. The mean teacher network refers to the temporal averaging of deep neural network parameters, which can produce a more accurate, stable representations than fixed weights at the end of training and is conventionally used for semi-supervised learning. Leveraging the success of the mean teacher framework in many studies, the proposed supervised learning method exploits the mean teacher network as an auxiliary model for better training of the main model, the student network. By learning the reliable intermediate representations derived from the mean teacher network as well as one-hot speaker labels, the student network is encouraged to explore more discriminative embedding spaces. The experimental results demonstrate that the proposed method relatively reduces the equal error rate by 11.61%, compared to the baseline system.

A GNSS anti-spoofing algorithm based on attention scheme in synchronization network

A GNSS anti-spoofing algorithm based on attention scheme jointly with label smooth loss was proposed and verified under designed GNSS dataset to detect potential and complex spoofing threat in synchronization network. Moreover the experimental results verified that the proposed algorithm showed some better performance of machine learning metrics comparing to conventional ways.

Deep learning-based virtual cytokeratin staining of gastric carcinomas to measure tumor–stroma ratio

The tumor–stroma ratio (TSR) determined by pathologists is subject to intra- and inter-observer variability. We aimed to develop a computational quantification method of TSR using deep learning-based virtual cytokeratin staining algorithms. Patients with 373 advanced (stage III [n = 171] and IV [n = 202]) gastric cancers were analyzed for TSR. Moderate agreement was observed, with a kappa value of 0.623, between deep learning metrics (dTSR) and visual measurement by pathologists (vTSR) and the area under the curve of receiver operating characteristic of 0.907. Moreover, dTSR was significantly associated with the overall survival of the patients (P = 0.0024). In conclusion, we developed a virtual cytokeratin staining and deep learning-based TSR measurement, which may aid in the diagnosis of TSR in gastric cancer.

Preprocessing approaches in machine learning-based groundwater potential mapping: an application to the Koulikoro and Bamako regions, Mali

Groundwater is crucial for domestic supplies in the Sahel, where the strategic importance of aquifers can only be expected to increase in the coming years due to climate change. Groundwater potential mapping is gaining recognition as a valuable tool to underpin water management practices in the region, and hence, to improve water access. This paper presents a machine learning method to map groundwater potential and illustrates it through an application to two regions of Mali. A set of explanatory variables for the presence of groundwater is developed first. Several scaling methods (standardization, normalization, maximum absolute value and min-max scaling) are used to avoid the pitfalls associated with the reclassification of explanatory variables. A number of supervised learning classifiers is then trained and tested on a large borehole database (n = 3,345) in order to find meaningful correlations between the presence or absence of groundwater and the explanatory variables. This process identifies noisy, collinear and counterproductive variables and excludes them from the input dataset. Tree-based algorithms, including the AdaBoost, Gradient Boosting, Random Forest, Decision Tree and Extra Trees classifiers were found to outperform other algorithms on a consistent basis (accuracy > 0.85), whereas maximum absolute value and standardization proved the most efficient methods to scale explanatory variables. Borehole flow rate data is used to calibrate the results beyond standard machine learning metrics, thus adding robustness to the predictions. The southern part of the study area was identified as the better groundwater prospect, which is consistent with the geological and climatic setting. From a methodological standpoint, the outcomes lead to three major conclusions: (1) because there is no aprioristic way to know which algorithm will work better on a given dataset, we advocate the use of a large number of machine learning classifiers, out of which the best are subsequently picked for ensembling; (2) standard machine learning metrics may be of limited value when appraising map outcomes, and should be complemented with hydrogeological indicators whenever possible; and (3) the scaling of the variables helps to minimize bias arising from expert judgement and maintains robust predictive capabilities.

DEEP LEARNING TISSUE ANALYSIS DIAGNOSES AND PREDICTS TREATMENT RESPONSE IN EOSINOPHILIC ESOPHAGITIS

Background: Eoinophilic Esophagitis (EoE) is a chronic inflammatory condition diagnosed by >=15 eosinophils (Eos) per high-power field (HPF). There is no gold standard for clinical remission and Eo-associated metrics are poorly correlated with symptoms. Deep learning can be used to explore the relationships of tissue features with clinical response. Objectives: To determine if deep learning can elucidate tissue patterns in EoE that predict treatments or symptoms at remission. Methods: We created two deep learning models using esophageal biopsies from histologically normal and EoE patients: one to identify Eos in esophageal biopsies and a second to broadly classify esophageal tissue as EoE vs. normal. We used these models to analyze biopsies at diagnosis and first remission timepoint, as defined by <15 Eos/HPF, in a subset of 19 treatment-naive patients. Differences in deep learning metrics between patient groups were assessed using Wilcoxon Rank-Sum tests. Results: All initial patients were symptomatic at diagnosis and a majority were still suffering from dysphagia at remission. The Eo identification model had a low mean (SD) error of -0.3 (11.5) Eos/HPF. Higher peak and average Eo counts at diagnosis were associated with higher likelihood of being on a food-elimination diet at remission than steroids or proton-pump inhibitor (p<0.05). The EoE classification model had an F1-score of 0.97 for distinguishing normal tissue from EoE. There was a significant decrease from diagnosis in the percentage of EoE-classified tissue among asymptomatic remission patients (p<0.05). Conclusions: Deep learning may have utility in diagnosing EoE and predicting future treatment response at diagnosis and resolution of symptoms at follow-up.

Multilabel CNN-Based Hybrid Learning Metric for Pedestrian Reidentification

Pedestrian reidentification has recently emerged as a hot topic that attains considerable attention since it can be applied to many potential applications in the surveillance system. However, high-accuracy pedestrian reidentification is a stimulating research problem because of variations in viewpoints, color, light, and other reasons. This work addresses the interferences and improves pedestrian reidentification accuracy by proposing two novel algorithms, pedestrian multilabel learning, and investigating hybrid learning metrics. First, unlike the existing models, we construct the identification framework using two subnetworks, namely, part detection subnetwork and feature extraction subnetwork, to obtain pedestrian attributes and low-level feature scores, respectively. Then, a hybrid learning metric that combines pedestrian attributes and low-level feature scores is proposed. Both low-level features and pedestrian attributes are utilized, thus enhancing the identification rate. Our simulation results on both datasets, i.e., CUHK03 and VIPeR, reveal that the identification rate is improved compared to the existing pedestrian reidentification methods.

Measuring human capital using global learning data

Human capital—that is, resources associated with the knowledge and skills of individuals—is a critical component of economic development1,2. Learning metrics that are comparable for countries globally are necessary to understand and track the formation of human capital. The increasing use of international achievement tests is an important step in this direction3. However, such tests are administered primarily in developed countries4, limiting our ability to analyse learning patterns in developing countries that may have the most to gain from the formation of human capital. Here we bridge this gap by constructing a globally comparable database of 164 countries from 2000 to 2017. The data represent 98% of the global population and developing economies comprise two-thirds of the included countries. Using this dataset, we show that global progress in learning—a priority Sustainable Development Goal—has been limited, despite increasing enrolment in primary and secondary education. Using an accounting exercise that includes a direct measure of schooling quality, we estimate that the role of human capital in explaining income differences across countries ranges from a fifth to half; this result has an intermediate position in the wide range of estimates provided in earlier papers in the literature5–13. Moreover, we show that average estimates mask considerable heterogeneity associated with income grouping across countries and regions. This heterogeneity highlights the importance of including countries at various stages of economic development when analysing the role of human capital in economic development. Finally, we show that our database provides a measure of human capital that is more closely associated with economic growth than current measures that are included in the Penn world tables version 9.014 and the human development index of the United Nations15.

SETTING, TRAINING AND TESTING OF A CONVOLUTIONAL NEURAL NETWORK TO COMLETE THE TASK OF THEMATIC PROCESSING OF SATELLITE IMAGES

We describe an algorithm based on a convolutional neural network that detects cloud and snow covers in satellite images. Algorithm accuracy was evaluated using machine learning metrics. The proposed algorithm is fully automatic