Data Selection Method Research Articles

Creating and Experimenting a New Parallel Corpus of English-Arabic Subtitles of Culinary Shows: A Useful Guide for Translating Food Across Cultures

This study presents and experiments a new English-Arabic corpus of food shows subtitles. It describes the data selection and extraction methods and suggests potential applications for the compiled corpus. It follows a corpus-driven approach to examine the strategies used to subtitle food-related terminology. The current study uses a new 259,956-word English-Arabic parallel corpus of some food shows, namely Nadiya Bakes, Nadiya Time to Eat, Somebody Feed Phil 1, and Somebody Feed Phil 2. The findings showed that Western culinary traditions, characterized by their accessibility and familiarity, exhibit a more seamless integration into the culinary vocabulary of Arabic. At the same time, the assimilation of Eastern gastronomies, renowned for their regional nuances and intricate flavors, appears to be comparatively less direct. Nevertheless, it is worth noting that Arabic, to a lesser degree, incorporates loanwords from a range of Eastern culinary traditions. The current piece of research provides some implications and recommendations for translators when handling culture-bound expressions, specifically food expressions. The practitioners are compelled to navigate a delicate equilibrium between applying loan translations and identifying fitting Arabic counterparts while considering the imperative factors of cultural sensitivity, intricate linguistic nuances, and contextual appropriateness.

Hydropedological digital mapping: machine learning applied to spectral VIS-IR and radiometric data dimensionality reduction

ABSTRACT Pedosphere-hydrosphere interface accounts for the association between soil hydrology and landscape, represented by topographic and Remote Sensing data support and integration. This study aimed to analyze different statistical radiometric and spectral data selection methods and dimensionality reduce environment-related data to support the classification of soil physical-hydric properties, such as soil basic infiltration rate (bir) and saturated hydraulic conductivity (Ksat); as well as to act in data mining processes applied to hydropedological properties digital mapping. Accordingly, research integrated information from Visible to Infrared (VIS-IR) spectral indices and Sentinel’s 2A mission Multispectral Instrument (MSI) sensor bands, terrain numerical modeling and aerogeophysics set to model soil-water content in two soil layers (0.00-0.20 m and 0.20-0.40 m). Pre-processed data were subjected to statistical analysis (multivariate and hypothesis tests); subsequently, the methods were applied (variation inflation factor - VIF, Stepwise Akaike information criterion – Stepwise AIC, and recursive feature elimination - RFE) to mine covariates used for Random Forest modeling. Based on the results, there were distinctions and singularities in spectral and radiometric data selection for each adopted method; the importance degree, and contribution of each one to soil physical-hydric properties have varied. According to the applied statistical metrics and decision-making criteria (highest R 2 and lowest RMSE / MAE), the chosen methods were RFE (0.00-0.20 m layers) and Stepwise AIC (0.20-0.40 m layers) - both concerned with the assessed variables (bir and Ksat). This approach captured the importance of environmental variables and highlighted their potential use in hydropedological digital mapping at Guapi-Macacu watershed.

Towards Better Generalization of Deep Neural Networks via Non-Typicality Sampling Scheme.

Improving the generalization performance of deep neural networks (DNNs) trained by minibatch stochastic gradient descent (SGD) has raised lots of concerns from deep learning practitioners. The standard simple random sampling (SRS) scheme used in minibatch SGD treats all training samples equally in gradient estimation. In this article, we study a new data selection method based on the intrinsic property of the training set to help DNNs have better generalization performance. Our theoretical analysis suggests that this new sampling scheme, called the nontypicality sampling scheme, boosts the generalization performance of DNNs through biasing the solution toward wider minima, under certain assumptions. We confirm our findings experimentally and show that more variants of minibatch SGD can also benefit from the new sampling scheme. Finally, we discuss an extension of the nontypicality sampling scheme that holds promise to enhance both generalization performance and convergence speed of minibatch SGD.

Effect of STEAM-Based Hybrid Based Learning Model on Students' Critical Thinking Skills

The purpose of the study was to determine the effect of STEAM-based hybrid based learning model on students' critical thinking skills. This research is a type of meta-analysis research. The research data came from 15 national and international journals. The process of searching for data sources through the google scholar database, Science Direct, Eric, Plos ONE and Wiley. The data selection method is the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) method. The inclusion criteria for this meta-analysis research are that the research must be published in the last 5 years 2018-2023, the research must use experimental or quasi-experimental methods, the research must have an experimental class that applies a hybrid-based learning model and a control class, the research comes from journals indexed by SINTA and Scopus, and the research has complete data to calculate the effect size. The data analysis technique is quantitative analysis by calculating the effect size value, research heterogeneity and publication bias with the help of the Comprehensive Meta-analysis (CMA) version.3 application. The results of the study concluded that the summary effect size or average effect size (p < 0.01; ES = 1.052) criteria were strong. This finding explains that the hybrid-based learning model has a significant effect on students' critical thinking skills. This STEAM-based hybrid-based learning model has a higher effect on students' critical thinking skills than the conventional model. Model has a high influence on students' critical thinking skills than conventional models

Exposure-response relation for vibration-induced white finger: effect of different methods for predicting prevalence.

A pooled analysis of vibration-induced white finger (VWF) in population groups of workers has been performed using the results of a published meta-analysis as source material (Nilsson T, Wahlström J, Burström L. Hand-arm vibration and the risk of vascular and neurological diseases a systematic review and meta-analysis. PLoS One. 2017:12(7):e0180795. https://doi.org/10.1371/journal.pone.0180795). The methods of data selection follow those described previously by Scholz et al. (in Scholz MF, Brammer AJ, Marburg S. Exposure-response relation for vibration-induced white finger: inferences from a published meta-analysis of population groups. Int Arch Occup Environ Health. 2023a:96(5):757-770. https://doi.org/10.1007/s00420-023-01965-w) to enable comparison with the results of the present work. The analyzed epidemiologic studies contain different prevalences of VWF observed after different durations of employment involving exposure to the vibration of power tools and machines. These prevalences are transformed to 10% prevalence by either linear or polynomial (i.e. "S"-shaped curvilinear) interpolation in order to compare with the exposure-response relation contained in the relevant international standard (ISO 5349-1:2001). An exposure-response relation is constructed using regression analysis for the time (in years) to reach 10% prevalence in a population group, when subjected to a daily vibration exposure calculated according to the procedures specified in the standard, A(8). Good fits to the data are obtained when polynomial and linear prevalence interpolation is used. The 95-percentile confidence intervals (CIs) of the exposure-response relation predicted by polynomial prevalence interpolation lie at somewhat larger lifetime exposures than those obtained by linear prevalence interpolation. Uncertainty in the precision of polynomial prevalence interpolation is mitigated by giving equal weight to linear interpolation when interpreting the results. When the 95-percentile CIs of the exposure-response models obtained by linear and polynomial prevalence interpolation are used to define the most probable exposure-response relation, the resulting common range of values includes the ISO exposure-response relation. It is proposed that an exposure-response relation for the onset of VWF derived from a regression analysis is specified in terms of the lower limit of its CI. Hence, when exposure measures are constructed according to the ISO standard and equal weight is given to the results of the 2 methods for interpolating prevalence described here, the ISO exposure-response relation would be considered to provide a conservative estimate for a 10% prevalence of VWF to develop in a population group, at least for A(8) > 4 m/s2. It thus remains the relation to use for assessing exposure to hand-transmitted vibration in the workplace. Additional research is needed to resolve inconsistencies in the ISO method for calculating daily exposures.

Effect of gas exchange data selection methods on resting metabolic rate estimation in young athletes.

This cross-sectional study analysed the effect of the gas exchange data selection methods on the resting metabolic rate (RMR) estimation and proposed a protocol shortening providing a suitable RMR estimation for young athletes. Sixty-six healthy young Brazilian athletes performed a 30-minute RMR assessment. Different methods of gas exchange data selection were applied: short and long-time intervals, steady-state (SSt), and filtering. A mixed one-way ANOVA was used to analyse the mean differences in gas exchange, RMR, respiratory exchange ratio (RER), and coefficients of variation across all methods. Additionally, paired Student's t-test were used to compare the first and best SSt RMR values for each SSt method (3, 4, and 5-min). The 5-min SSt method provided the lowest RMR estimate (1454 kcal.day-1). There was a statistical difference between methods (F = 2.607, p = 0.04), but they presented a clinically irrelevant absolute difference (~36 kcal.day-1). There were no differences in RER among methods. In addition, using the SSt method, 12 minutes of assessment were enough to obtain a valid estimation of RMR. The 5-min SSt method should be employed for assessing the RMR among young athletes, considering the possibility of obtaining a shortened assessment (~12 min) with an acceptable and low coefficient of variation.

A rapid outliers detection and correction method for external ballistic velocity measurement data

Due to the influence of various sudden and abnormal factors during the tracking environment, equipment working conditions, and operation process, there are inevitable outliers in the tracking measurement process of aircraft such as carrier rockets, artificial satellites, and missiles. Therefore, the prerequisite for ensuring the reliability of processing results is to timely and accurately detect and correct outliers. In this paper, we proposed the sliding window-based variable degree B-spline function method (SWVD B-spline), which can handle isolated outliers and spotted outliers. SWVD B-spline uses variable degree B-spline function to model observed data in sliding windows, which can detect and correct outliers point by point along with window sliding. Then, we propose an initial window data selection method to remove outliers in initial windows to ensure the processing effect. In addition, because there are often inflection points in external ballistic velocity measurement data, differential evolution is used to optimize variable degree B-spline in windows that include inflection points to improve processing accuracy. The experimental results verify that SWVD B-spline can handle various outliers rapidly and efficiently.

Impact of methods for data selection on the day-to-day reproducibility of resting metabolic rate assessed with four different metabolic carts

Background and aimsAccomplishing a high day-to-day reproducibility is important to detect changes in resting metabolic rate (RMR) and respiratory exchange ratio (RER) that may be produced after an intervention or for monitoring patients' metabolism over time. We aimed to analyze: (i) the influence of different methods for selecting indirect calorimetry data on RMR and RER assessments; and, (ii) whether these methods influence RMR and RER day-to-day reproducibility. Methods and resultsTwenty-eight young adults accomplished 4 consecutive RMR assessments (30-min each), using the Q-NRG (Cosmed, Rome, Italy), the Vyntus CPX (Jaeger-CareFusion, Höchberg, Germany), the Omnical (Maastricht Instruments, Maastricht, The Netherlands), and the Ultima CardiO2 (Medgraphics Corporation, St. Paul, Minnesota, USA) carts, on 2 consecutive mornings. Three types of methods were used: (i) short (periods of 5 consecutive minutes; 6–10, 11–15, 16–20, 21–25, and 26–30 min) and long time intervals (TI) methods (6–25 and 6–30 min); (ii) steady state (SSt methods); and, (iii) methods filtering the data by thresholding from the mean RMR (filtering methods). RMR and RER were similar when using different methods (except RMR for the Vyntus and RER for the Q-NRG). Conversely, using different methods impacted RMR (all P ≤ 0.037) and/or RER (P ≤ 0.009) day-to-day reproducibility in all carts. The 6–25 min and the 6–30 min long TI methods yielded more reproducible measurements for all metabolic carts. ConclusionThe 6–25 min and 6–30 min should be the preferred methods for selecting data, as they result in the highest day-to-day reproducibility of RMR and RER assessments.

A Study on Factors Associated with Child Sexual Abuse and Recognizing the Severity: Special Reference to Galle District

Child Sexual Abuse has been a global epidemic with devastating consequences. One in four girls and one in six boys have been experienced some form of sexual abuse in their tender age in the world. According to Police statistics, Child Sexual Abuse (CSA) cases is growing in recent years in Sri Lanka too. Galle is among the four districts where the reported child abuse cases high and the reported CSA complaints are increasing extraordinarily. Also, there is no previous research have been done in the Southern part of the country regarding the crisis of CSA. So, main objective of this study is to determine the key risk factors that affected to a CSA in Galle Police Division, and to develop suitable regression and machine learning models to predict the severity of CSA. 225 CSA cases reported to Police Child and Women Bureau of Galle Police Division during the period 2017 – 2020 were treated for this study. Out of twenty-one risk which were found from literature and knowledge of domain experts, sixteen variables showed a significant relationship with response variable severity of CSA according to chi-square test of association. Traditional OLR model was performed to predict severity of CSA and to detect key risk factors to a CSA with two different data selection methods. Next, machine learning techniques: Decision Tree, SVM, and PNN were trained to classify severity of CSA. Random over-sampling technique was used to overcome the class imbalanced problem persists in the dataset. Finally, bagging technique was executed to conserve robustness of models and to improve performance. The OLR model classified the severity of CSA with 68.85% accuracy. Machine learning techniques, Decision Tree, SVM and PNN model classified the severity of CSA with an accuracy of 82.15%, 77.68% and 85.25% respectively. PNN model performed with higher accuracy better than other fitted models. The results obtained from this study can be used to take precautions and to arrange awareness sessions for adults to reduce CSA in Galle Police Division. Also, the study can be extended to the whole island to reduce CSA and to make it a better place for children.

Iterative pseudo-labelling with SoftMax probability in text classification

Semi-supervised learning is one of the potential research fields in text classification. In this paper, semi-supervised pseudo-label training experiments are conducted using the BERT model that has been pre-trained as a baseline. Only 20% of the original dataset is used for the new training set after segmenting the training set. The raw corpus used for pseudo-label training consists of the remaining 80% of data after labels are removed, while the original test set is still utilized. The results indicate that the key to the semi-supervised pseudo-labelling method is the performance of the original model and reasonable data filtering techniques. Even though the SoftMax value used for data filtering is not precisely equivalent to model prediction accuracy, experimental results show it can somewhat reduce the error propagation problem of the model. This is consistent with earlier research. However, using SoftMax as the threshold for data screening can't bring enough benefits to the model training and make it surpass the training performance of the original data set. As a result, future studies will focus on improving the accuracy of pseudo-labelling with a more suitable data selection method to better the model's performance.

Psi-entific Approach to Post-Mortem Survival: Employing the Multiple Sources of Psi (MSoP) and Discarnate Psi Hypotheses in the Calculation of a Drake-S Equation

The continued investigations of the paranormal facilitated the development of novel theories and methodologies. Psi functioning of the living and the deceased in survival phenomena suggested the living agent psi (LAP) and discarnate psi hypotheses, but neither has demonstrated sufficient explanatory power to claim superiority in explaining survival data. Mediumship studies cannot determine whether paranormal information is sourced by means of discarnate psi or LAP, presenting the source-of-psi problem. Anomalous information can be obtained from joint sources (LAP, survival, or some other source), which supports the multiple sources of psi (MSoP) hypothesis. The maximized explanatory potential of the MSoP hypothesis makes the inclusion of the LAP and discarnate psi factors in the calculation of a Drake-S equation for post-mortem survival required and appropriate. This paper concludes that 1) the aggregate effect of skeptical explanations for survival was calculated at 65.6%, leaving 35.4% to paranormal explanations, which contradicts skeptical claims and is inconsistent with the existing laws of conventional science; 2) 16% of paranormal experiences reported among the general population appeared genuine; and 3) the calculated purified probability for all paranormal phenomena equaling 40% can be attributed to paranormal causes. This suggests reasonable plausibility of the survival hypothesis. To refine the existing factors and find new empirical factors related to known confounds and anomalous effects, future research should include more robust procedures and methods of data selection, gathering, and analysis.

Modified DEMATEL Method Based on Objective Data Grey Relational Analysis for Time Series

Smart data selection can quickly sieve valuable information from initial data. Doing so improves the efficiency of analyzing situations to aid in better decision-making. Past methods have mostly been based on expert experience, which may be subjective and inefficient when dealing with large, complex datasets. Recently, the system analysis method has been exploited to find the key data. However, few studies address the indirect effects and heterogeneity of time series data. In this study, a data selection method, the modified Decision-Making Trial and Evaluation Laboratory (DEMATEL) method based on the objective data grey relational analysis (GRA), is used to enhance the ability to analyze time-series data. GRA was first applied to assess the direct impact in the raw data indicators. Then, a modified DEMATEL was adopted to find the overall impact by including the indirect impact and data heterogeneity. We applied the method to analyze the Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) dataset and perform the remaining useful life (RUL) prediction of aircraft engines. The results suggest that our method predicts well. Our work offers a nuanced approach of identifying key information in time series data and has potential applications.

System-Information and Cognitive Technologies of Man-Made Infrastructure Cyber Security

The complication of technological processes due to the modernization of complex technological processes aggregates with a distributed spatial infrastructure requires the use of new control systems and computer and information network technologies. Accordingly, this poses the problem of revising the basic concepts of information-measuring systems building, developing software and hardware for the implementation of the target management process. This requires development of new approaches to presentation, processing, and display of data about the aggregates state of energy-active objects and the entire information structure. For management under threat conditions, it is necessary to take into account the features of the information infrastructure, data selection and processing methods, methods and algorithms for classifying the situation, which are formed from blocks of data obtained from each unit and the technological process as a whole. Information about the current state of system and infrastructure is necessary for the formation, adoption and implementation of management decisions which is the basis for coordination strategies synthesis. Appropriate target orientation, reasonable indicators of real process trajectories divergence from the target state determine the probability of object attack. Knowledge and decision-making procedures for the coordination of managerial actions is based on the strategic target orientation of the structure, their professionalism and the level of intellectual, cognitive and scientific training which is the basis for correct situation interpretation of countermeasures against threats formation. In the event of active threats complex on man-made systems in a certain region (resource, cognitive, system, information) and natural disasters or military operations, the threats lead to active destruction or failure of the production process. In order to functionally withstand related production structures, when loyal to the industrial relations concept, they need to integrate at the strategic management level on common goal basis to reduce risks. If necessary, to increase sustainability, they can be integrated at operational management level using interconnections at the production and resource levels. To do this, it is necessary to develop a strategic and goal-oriented management system behaviour line, which requires informational and intelligent data processing at the highest level using cognitive creative methods. For each level of the infrastructure hierarchy, oriented towards strategic goals in the global infrastructure dynamic environment, methods of assessing the situation to detect failures and the actions of attacks have been developed, based on which countermeasures are formed depending on the type of threats.

A data selection method for matrix effects and uncertainty reduction for laser-induced breakdown spectroscopy

Severe matrix effects and high signal uncertainty are two key bottlenecks for the quantitative performance and wide applications of laser-induced breakdown spectroscopy (LIBS). Based on the understanding that the superposition of both matrix effects and signal uncertainty directly affects plasma parameters and further influences spectral intensity and LIBS quantification performance, a data selection method based on plasma temperature matching (DSPTM) was proposed to reduce both matrix effects and signal uncertainty. By selecting spectra with smaller plasma temperature differences for all samples, the proposed method was able to build up the quantification model to rely more on spectra with smaller matrix effects and signal uncertainty, therefore improving final quantification performance. When applied to quantitative analysis of the zinc content in brass alloys, it was found that both accuracy and precision were improved using either a univariate model or multiple linear regression (MLR). More specifically, for the univariate model, the root-mean-square error of prediction (RMSEP), the determination coefficients (R 2) and relative standard derivation (RSD) were improved from 3.30%, 0.864 and 18.8% to 1.06%, 0.986 and 13.5%, respectively; while for MLR, RMSEP, R 2 and RSD were improved from 3.22%, 0.871 and 26.2% to 1.07%, 0.986 and 17.4%, respectively. These results prove that DSPTM can be used as an effective method to reduce matrix effects and improve repeatability by selecting reliable data.

Everywhere: A Framework for Ubiquitous Indoor Localization

Smartphones have become an integral part of daily human life and enable almost unlimited coverage of human mobility. Thus, collecting pervasive crowdsourced signatures is feasible. Autonomous localization of such signatures promotes the development of a self-deployable and ubiquitous indoor positioning system (IPS). However, previous crowdsourcing-based IPSs have not considered leveraging such data for developing ubiquitous IPSs. They have relied on methods for data selection and sources for localization adjustment that could work against realizing a ubiquitous system. In contrast, this study introduces a framework Everywhere that leverages crowdsourced data to develop a ubiquitous IPS and addresses existing challenges while developing such systems. Particularly, inertial data selection criteria are proposed to autonomously generate traces with better localization. Moreover, pervasive global navigation satellite system (GNSS) data are leveraged to adjust trace localization, while simultaneously introducing a deploying location (inside elevators) of one anchor node. The node surveys all the floors while reducing the localization error, especially for the buildings surrounded by GNSS-denied areas. Additionally, cumulative data densification is leveraged to realize pervasive resources within the building, thereby boosting trace adjustment and extending database spatial coverage. Furthermore, a better selection of neighboring fingerprints is proposed to enhance online fingerprinting. Such a framework can promote a ubiquitous IPS development for buildings regardless of whether they are surrounded by open sky or GNSS-denied areas.

User data selection method using received signal strength indicator for semi-supervised learning in fingerprint localization

The fingerprint technique is used as an indoor localization method. This study uses a CNN-based indoor fingerprint localization method. The estimation accuracy of CNN improves as the number of AP information (AP identifiers and received signal strength indicator) increases. However, gathering AP information is time-consuming and costly. The problem can be solved using UD (AP information users measured). However, the UD measuring method does not know the user’s exact location. Therefore, it is essential to choose UD that is accurately estimate and use it for CNN training. In this study, we propose a method for selecting UDs that makes use of the RSSI similarity between AP information and UD.

Práticas Integrativas e Complementares na atuação dos enfermeiros da Atenção Primária à Saúde

Abstract Objective to analyze the knowledge and use of integrative and complementary practices in the work of nurses in Primary Health Care. Method a cross-sectional study carried out in 85 Basic Health Units in Curitiba-PR in 2021. The sample consisted of 195 recruits recruited using the Snowball technique. The were confirmed by means of an electronic data selection instrument and methods by electronic descriptive data methods. Results 43 (22.1%) of the participants use integrative and complementary practices (PIC) in care, and 31 (15.9%) reported having training in some modality. The use of 17 PIC modalities was identified, 33 (47.8%) in auriculotherapy and 7 (10.1%) in reiki. The Ranking item on the theoretical knowledge of the average participants on the topic was >3.89, demonstrating agreement with the statements. About how participants experience the PIC, due to the recognition of nursing practices. Conclusion the study contributes to the characterization of nurses' performance in primary care based on integrative and complementary practices. The findings serve as a subsidy for professionals and managers to identify gaps in knowledge, in action, as well as in the management of primary care health services in order to strengthen the professional autonomy of those who use these practices in care.

Data selection method for restoring a tsunami source form

The reconstruction of a tsunami source as a solution to the inverse problem in mathematical physics relies on the use of the truncated singular value decomposition method (a variant of the least squares method) for inverting remote records of the tsunami wave. The proposed method allows one to overcome the inevitable instability of the numerical solution. The result of inversion depends on the choice of the observation system, actual bathymetry and data noise level. Within the developed approach, a methodology for choosing key inversion options and an optimal dataset which provide the best accuracy of a tsunami source recovery is discussed. It is based on analyzing the distribution of the specific energy generated by all spatial modes at the locations of the active sensors. The peculiarity of the algorithm is that the use of the most informative data allows one without re-computation of the direct problem to obtain wave amplitudes at the points of interest (let us call them as “fictitious” stations) where there were no observations, but those that were considered in preliminary calculations. Three real-life events, the Chilean Illapel tsunami on September 16, 2015, the tsunami near the Solomon Islands on February 6, 2013, and the Shikotan tsunami on October 5, 1994, are used as examples of the proposed approach. The results obtained allow one hope for using of this approach in practice.

Application of Discriminative Training Algorithm Based on Intelligent Computing in English Translation Evaluation

Abstract Aiming at the higher correlation between the objective evaluation of computer English speech and the subjective evaluation of experts, an acoustic model based on discriminative training is proposed to improve the confidence score of objective evaluation. First, the process of obtaining the pronunciation quality evaluation score of the speech vector by the forced matching algorithm is introduced, and then the mathematical theory of hypothesis testing is used to prove that the acoustic model trained based on the discriminative algorithm ‘minimum phoneme error’ is more effective than the acoustic model based on the traditional maximum likelihood algorithm. Confidence scores close to subjective assessments are obtained. By calculating the correlation coefficient of the subjective and objective evaluation results, the experiment verifies that the speech evaluation system using the discriminative acoustic model can give a higher confidence score and proposes a data selection method based on dynamic weighting, which is applied to continuous speech recognition in the discriminative training of the acoustic model. This method combines the posterior probability and the phoneme accuracy rate to select the data. First, the Beam algorithm of the posterior probability is used to trim the word graph. On this basis, according to the error rate of the candidate path where the candidate word is located, the probability dynamically assigns different weights to the candidate words; second, by calculating the confusion degree between the phoneme pairs, different penalty weights is dynamically added to the easily confused phoneme pairs to calculate the phoneme accuracy; finally, the expected accuracy of the obtained arc is calculated on the basis of the probability distribution. The Gaussian function is used to softly weight the expected phoneme accuracy of all competing arcs. The experimental results show that compared with the minimum phoneme error criterion, the dynamic weighting method has higher recognition accuracy and can effectively reduce the training time.

Drill the Cork of Information Bottleneck by Inputting the Most Important Data.

Deep learning has become the most powerful machine learning tool in the last decade. However, how to efficiently train deep neural networks remains to be thoroughly solved. The widely used minibatch stochastic gradient descent (SGD) still needs to be accelerated. As a promising tool to better understand the learning dynamic of minibatch SGD, the information bottleneck (IB) theory claims that the optimization process consists of an initial fitting phase and the following compression phase. Based on this principle, we further study typicality sampling, an efficient data selection method, and propose a new explanation of how it helps accelerate the training process of the deep networks. We show that the fitting phase depicted in the IB theory will be boosted with a high signal-to-noise ratio of gradient approximation if the typicality sampling is appropriately adopted. Furthermore, this finding also implies that the prior information of the training set is critical to the optimization process, and the better use of the most important data can help the information flow through the bottleneck faster. Both theoretical analysis and experimental results on synthetic and real-world datasets demonstrate our conclusions.