Variable Representation Research Articles

Causality-embedded reconstruction network for high-resolution fault identification in chemical process

Fault identification is essential for analyzing the root causes and propagation of faults. Traditional identification based on contribution plots often suffer from the smearing effect, a phenomenon where the residuals of fault variables smearing over other non-fault process variables due to the complex interconnections within the processes, potentially leading to ambiguous or inaccurate diagnosis. To address this, we propose the Causality-Embedded Reconstruction Network (CERN) based high-resolution fault identification for chemical process. CERN is an advanced graph network model that embeds causality, featuring a dual-decoding scheme comprising a causal decoder and an attribute decoder. This design ensures that feature reconstruction is guided by variables with established causal relationships, effectively isolating the feature representations of normal variables from the influence of fault variables that lack a causal connection. By leveraging this causality-embedded reconstruction, the smearing effect can be alleviated. Experiments conducted on a Continuous Stirred-Tank Reactor system and a real-world Continuous Catalytic Reforming process to demonstrate that the proposed CERN-based fault identification method not only improves the accuracy in identifying fault variables but also offers explanatory insights and traceability for fault diagnosis, showing prospects for industrial process safety.

Sociodemographic Correlates of Postpartum Depression: A Survey-Based Study.

Objective: Post-partum depression (PPD) has been reported in about one-seventh to one-tenth of women. The aim of this study is to identify the demographic, obstetrics, social, and psychological risk factors of PPD among the eastern region of the United Arab Emirates. Method : A community-based cross-sectional study was performed on 200 women who had a recent singleton pregnancy and delivered newborn within past six months via convenience sampling and email snowballing. Several demographics, obstetrics, social, and psychological factors of the respondents were assessed using a survey form. The Edinburgh depression rating scale (EDRS) was used for the identification of women with possible PPD. Descriptive statistics were utilized for the representation of demographic variables, whereas Chi-square test was employed to assess categorical variables. Also, logistic regression was applied to evaluate the association of investigated variables and PPD. Results: The median EDRS score amongst the study participants was found to be 11 (0-26). The prevalence of PPD was found to be 57% in the studied population. Significant differences were observed in the adverse life events, emotional supports, marital conflicts and history of depression of the participants with and without PPD (P < 0.01). The risk factors significantly associated with PPD were age of the newborn (OR = 6.50, 95%CI: 1.17-19.91), marital relationship (OR = 4.15, 95%CI: 1.31-15.22), maternal educational level (OR = 5.10, 95%CI: 4.30-16.58), adverse life events (OR = 9.32, 95%CI: 1.33-35.32), and history of depression (OR = 5.24, 95%CI: 3.14-11.96). Conclusion: Given the findings, there is an urgent need for policy initiatives to address the identified risk factors, such as improving access to education, strengthening supportive marital relationships, and providing comprehensive mental health services for pregnant women.

Impacts of physical parameterization scheme of snow and soil properties over the Tibetan Plateau on the simulation of summer rainfall in China

Accurate representation of soil and snow state variables in the land surface model is crucial for good simulation results. In order to investigate the potential impacts of the overestimated snow cover over the Tibetan Plateau (TP) on precipitation over China, we designed comparative experiments over TP with and without updated soil texture data, soil hydrological parameters, and the fresh snow density scheme to address this issue with the Weather Research and Forecasting Model (WRF). The results show that snow cover fraction (SCF) and surface albedo in winter on the TP were overestimated in the WRF control experiment (CTL), with albedo bias exceeding 0.4 in the western TP. The SCF in the WRF improved experiment (WRFIM) is about 8% lower than that of the CTL, and the bias is decreased to 0.2–0.3 in the interior of the TP. The thermal and dynamic effects of the TP in winter have a significant impact on the weather and climate of East Asia during summer. Thus, the simulated summer rainfall bias is significantly reduced by WRFIM in North China, and the mid-lower reaches of the Yangtze River, along with the spatial correlation coefficient between the observed-based and WRF-simulated rainfall increased from 0.76 in CTL to 0.88 in WRFIM. The WRFIM-simulated reductions in snow cover and albedo cause a decrease in upward shortwave radiation and eventually increase net surface radiation. The horizontal temperature gradient of the atmosphere is increased in the WRFIM compared to the CTL, and the increased temperature in the lower and middle layers accommodates more water vapor, which is not conducive to precipitation formation. Correspondingly, the Western Pacific Subtropical High at 500 hPa in the middle troposphere weakened. The low-level anticyclonic circulation at 850 hPa also weakened to the east, affecting water vapor transport in eastern China. It leads to a decrease in the water vapor transported northward by southeasterly from the western Pacific Ocean and a decrease in the water vapor flux reaching North China. This work indicates that accurate characterization of soil properties and snow cover on the TP is crucial for improving the simulation bias of summer rainfall in China.

Heterogeneous Graph Neural Network with Personalized and Adaptive Diversity for News Recommendation

The emergence of online media has facilitated the dissemination of news, but has also introduced the problem of information overload. To address this issue, providing users with accurate and diverse news recommendations has become increasingly important. News possesses rich and heterogeneous content, and the factors that attract users to news reading are varied. Consequently, accurate news recommendation requires modeling of both the heterogeneous content of news and the heterogeneous user-news relationships. Furthermore, users’ news consumption is highly dynamic, which is reflected in the differences in topic concentration among different users and in the real-time changes in user interests. To this end, we propose a Heterogeneous Graph Neural Network with Personalized and Adaptive Diversity for News Recommendation (DivHGNN). DivHGNN first represents the heterogeneous content of news and the heterogeneous user-news relationships as an attributed heterogeneous graph. Then, through a heterogeneous node content adapter, it models the heterogeneous node attributes into aligned and fused node representations. With the proposed attributed heterogeneous graph neural network, DivHGNN integrates the heterogeneous relationships to enhance node representation for accurate news recommendations. We also discuss relation pruning, model deployment, and cold-start issues to further improve model efficiency. In terms of diversity, DivHGNN simultaneously models the variance of nodes through variational representation learning for providing personalized diversity. Additionally, a time-continuous exponentially decaying distribution cache is proposed to model the temporal dynamics of user real-time interests for providing adaptive diversity. Extensive experiments on real-world news datasets demonstrate the effectiveness of the proposed method.

Comparative analysis of indices in capturing the onset and withdrawal of the South Asian Summer Monsoon

The South Asian summer monsoon (SASM) exerts a profound influence on South Asia and the southern Tibetan Plateau. The timing of its onset and withdrawal significantly impacts regional rainfall, a critical water source for agriculture and the economy. Various SASM indices, employing different variables, have been employed to gauge monsoon onset and withdrawal, each demonstrating distinct characteristics. This study provides a comprehensive comparison of prominent SASM indices related to onset and withdrawal, revealing similar timing but varying magnitudes. Across nearly all indices, a consistent pattern emerges, indicating a trend towards earlier onset and delayed retreat during 1979–2018, marked by pronounced interdecadal variability, with a notable turning point around 1997. The earlier onset and later retreat are likely to enhance rainfall and potential water resources for South Asia and the Tibetan Plateau. Delving into the mechanisms revealed a delayed connection between the SASM onset to the large-scale sea surface temperature (SST) anomalies characterized by the Oceanic Niño Index (ONI) in the preceding spring, but a simultaneous connection between SASM withdrawl and ONI in autumn. Current index definitions, relying on single or dual variables, may fall short in accurately depicting monsoon onset and withdrawal. To address this, we introduce a novel monsoon index derived from multiple large-scale circulation variables, offering improved trend capture and enhanced representation of interannual variability in SASM onset and withdrawal. This study advances our understanding of SASM indices and their correlation with monsoon rainfall, providing insights into the dynamics of SASM onset and withdrawal.

Predicting Student Placement in College Using Machine Learning

Abstract: The research on student placement prediction in higher education has been a focal point, addressing challenges related to balanced accuracy and generalization across diverse datasets. Prior studies grappled with issues in feature representation and interpretability. This study aims to overcome these challenges by introducing a comprehensive machine learning framework for student placement prediction, leveraging advanced techniques in exploratory data analysis, preprocessing, and model evaluation. Drawing on previous research experiences, our proposed work targets specific issues related to feature engineering, categorical variable representation, and result interpretability. The methodology employs key libraries, including NumPy, pandas, Matplotlib, Seaborn, Plotly, scikit-learn, WordCloud, and DateTime, for efficient data manipulation, visualization, and analysis. Ensemble learning techniques, such as Random Forest and XGBoost, along with traditional algorithms like Decision Trees and K-Nearest Neighbors, contribute to enhancing predictive accuracy and model robustness. To fine-tune the models, a Randomized Search for Hyperparameters is implemented for the XGBoost classifier, optimizing parameters like learning rate, maximum depth, minimum child weight, gamma, and colsample by tree. This approach effectively addresses overfitting and underfitting issues, maximizing overall model performance. The accuracy percentages achieved through our models represent significant advancements. For instance, the Decision Tree model achieves an accuracy of 87.74%, the Random Forest model achieves 87.60%, the XGBoost model achieves 87.60%, and the K-Nearest Neighbors model achieves 85.18%. These results underscore the effectiveness of our approach in achieving high accuracy while maintaining interpretability.

Multi-decadal pacemaker simulations with an intermediate-complexity climate model

Abstract. In this paper, we first describe the main features of a new version of the International Centre for Theoretical Physics global atmospheric model (SPEEDY) with improved simulation of surface fluxes and the formulation of a three-layer thermodynamic ocean model (TOM3) suitable to explore the coupled extratropical response to tropical ocean variability. Then, we present results on the atmospheric model climatology, highlighting the impact of the modifications introduced in the model code, and show how important features of interdecadal and interannual variability are simulated in a “pacemaker” coupled ensemble of 70-year runs, where portions of the tropical Indo-Pacific are constrained to follow the observed variability. Despite the very basic representation of variations in greenhouse forcing and heat transport to the deep ocean (below the 300 m domain of the TOM3 model), the coupled ensemble reproduces the variations in surface temperature over land and sea with very good accuracy, confirming the role of the Indo-Pacific as a “pacemaker” for the natural fluctuations of global-mean surface temperatures found in earlier studies. Atmospheric zonal-mean temperature trends over 50 years are also realistically simulated in the extratropical lower troposphere and up to 100 hPa in the tropics. On the interannual scale, sea-surface temperature (SST) variability in sub-tropical and tropical regions not affected by SST relaxation is underestimated (mostly because of the absence of dynamically induced variability), while extratropical SST variability during the cold seasons is comparable to that observed. Atmospheric teleconnection patterns and their connections with SST are reproduced with high fidelity, although with local differences in the amplitude of regional features (such as a larger-than-observed response of extratropical SST to North Atlantic Oscillation variability). The SPEEDY-TOM3 model also reproduces the observed connection between averages of surface heat fluxes over the oceans and land surface air temperature in the wintertime northern extratropics. Overall, as in earlier versions of SPEEDY, the fidelity of the simulations (both in terms of climatological means and variability) is higher near the surface and in the lower troposphere, while the negative impacts of the coarse vertical resolution and simplified parameterizations are mostly felt in the stratosphere. However, the improved simulation of surface heat fluxes and their impact on extratropical SST variability in this model version (obtained at a very modest computational cost) make the SPEEDY-TOM3 model a suitable tool to investigate the coupled response of the extratropical circulation to interannual and inter-decadal changes of tropical SST in ensemble experiments.

Robust reduced-order machine learning modeling of high-dimensional nonlinear processes using noisy data

Autoencoder-based reduced-order machine learning models have been developed for modeling and predictive control of nonlinear chemical processes with high dimensionality such as discretization of reaction–diffusion processes. However, in the presence of data noise, autoencoders may over-fit the training data and subsequently learn an inaccurate low-dimensional representation of the process variables. This leads to an inaccurate prediction model when the models are integrated with model predictive control (MPC). To address this issue, this work develops a novel machine-learning-based reduced-order modeling method by integrating SpectralDense layers into autoencoders and incorporating them with recurrent neural networks. We demonstrate that the new architecture of autoencoders using SpectralDense layers is more robust against over-fitting than conventional autoencoders in the presence of data noise, which improves the prediction accuracy in MPC. A diffusion–reaction process simulation example is used to demonstrate that the robust autoencoders outperform those using conventional layers for reduced-order modeling in predictive control.

Image‐based characterization of flocculation processes through PLS inspired representation learning in convolutional neural networks

AbstractMonitoring of flocculation processes such as those used in downstream processing of a fermentation broth is essential for process control. One approach is to apply microscopic imaging combined with image analysis for characterizing the state of the process. In this work, we investigate and compare the use of supervised feedforward convolutional neural network (CNN) architectures to predict the process states from the image information and compare the results with the traditional alternative of characterizing flocs based on manually engineered image features guided by human expertise. From a well‐defined image data set representing six process states, the objective is to establish end‐to‐end classification models which are accurate but at the same time learn meaningful latent variable space representations. Specifically, we evaluate three different CNN architectures with varying degrees of regularization and compare results with logistic regression models based on inputs from two different traditional feature engineering methods. By applying global average pooling as a structural regularizer to the CNN architecture, we significantly improve the generalization performance in comparison with the classification accuracies of the traditional feature engineered models. Furthermore, we show that by imposing a projection to latent structures (PLS) like regularization framework onto the CNN, it can also learn a latent variable representation that mimics the features selected by human expertise.

Adherence to the International Committee of Medical Journal Editors–recommended gender equity policy in nursing journals listed in MEDLINE or PubMed Central: a descriptive study

Purpose: The evolving landscape of nursing research emphasizes inclusive representation. The International Committee of Medical Journal Editors (ICMJE) has established guidelines to ensure the fair representation of various demographic variables, including age, sex, and ethnicity. This study aimed to evaluate the adherence of nursing journals indexed in MEDLINE or PubMed Central to the ICMJE’s directives on gender equity, given that journals indexed in MEDLINE and PubMed Central typically adhere to the ICMJE’s guidelines.Methods: A descriptive literature review methodology was employed to analyze 160 nursing journals listed in two databases as of July 28, 2023. The website of each journal was searched, and the most recent original article from each was selected. These articles were then evaluated for their alignment with the ICMJE guidelines on gender equity. Descriptive statistics were applied to categorize and enumerate the cases.Results: Of the articles reviewed from 160 journals, 115 dealt with human populations. Of these, 93 required a description of gender equity. Within this subset, 83 articles distinguished between the genders of human subjects. Gender-based interpretations were provided in 15 articles, while another 68 did not offer an interpretation of differences by gender. Among the 10 articles that did not delineate gender, only two provided a rationale for this omission.Conclusion: Among recent articles published in the nursing journals indexed in MEDLINE and PubMed Central, only 16.1% presented clear gender analyses. These findings highlight the need for editors to strengthen their dedication to gender equity within their editorial policies.

Bayesian Semiparametric Longitudinal Inverse-Probit Mixed Models for Category Learning.

Understanding how the adult human brain learns novel categories is an important problem in neuroscience. Drift-diffusion models are popular in such contexts for their ability to mimic the underlying neural mechanisms. One such model for gradual longitudinal learning was recently developed in Paulon et al. (J Am Stat Assoc 116:1114-1127, 2021). In practice, category response accuracies are often the only reliable measure recorded by behavioral scientists to describe human learning. Category response accuracies are, however, often the only reliable measure recorded by behavioral scientists to describe human learning. To our knowledge, however, drift-diffusion models for such scenarios have never been considered in the literature before. To address this gap, in this article, we build carefully on Paulon et al. (J Am Stat Assoc 116:1114-1127, 2021), but now with latent response times integrated out, to derive a novel biologically interpretable class of 'inverse-probit' categorical probability models for observed categories alone. However, this new marginal model presents significant identifiability and inferential challenges not encountered originally for the joint model in Paulon et al. (J Am Stat Assoc 116:1114-1127, 2021). We address these new challenges using a novel projection-based approach with a symmetry-preserving identifiability constraint that allows us to work with conjugate priors in an unconstrained space. We adapt the model for group and individual-level inference in longitudinal settings. Building again on the model's latent variable representation, we design an efficient Markov chain Monte Carlo algorithm for posterior computation. We evaluate the empirical performance of the method through simulation experiments. The practical efficacy of the method is illustrated in applications to longitudinal tone learning studies.

Dual-attention LSTM autoencoder for fault detection in industrial complex dynamic processes

Complex dynamic characteristics resulting from multi-system coupling and closed-loop control are ubiquitous in modern industrial process data, presenting significant challenges for process fault detection. However, conventional data-driven fault detection methods assume the data to be static or slightly dynamic. Addressing the complex dynamic characteristics and nonlinearity inherent in industrial processes, this paper proposes a dual-attention long short-term memory autoencoder (DALSTM-AE) for fault detection in dynamic processes. Long short-term memory (LSTM) and autoencoder (AE) are combined into a special encoder-decoder LSTM architecture to learn both dynamic features and deep representations of variables in an unsupervised manner. Then, a dual-attention module is embedded in the decoder to properly learn the temporal dependencies associated with long input sequences and retain the most critical information. In addition, based on the reconstruction results of the DALSTM-AE model, two monitoring statistics are designed for fault detection. Finally, the effectiveness and superiority of the proposed method are fully demonstrated through case studies on a numerical simulation example, the Tennessee Eastman (TE) benchmark process, and practical coal pulverizing systems in power plants.

Development of a universal method for vibrational analysis of the terminal alkyne C≡C stretch.

The terminal alkyne C≡C stretch has a large Raman scattering cross section in the "silent" region for biomolecules. This has led to many Raman tag and probe studies using this moiety to study biomolecular systems. A computational investigation of these systems is vital to aid in the interpretation of these results. In this work, we develop a method for computing terminal alkyne vibrational frequencies and isotropic transition polarizabilities that can easily and accurately be applied to any terminal alkyne molecule. We apply the discrete variable representation method to a localized version of the C≡C stretch normal mode. The errors of (1) vibrational localization to the terminal alkyne moiety, (2) anharmonic normal mode isolation, and (3) discretization of the Born-Oppenheimer potential energy surface are quantified and found to be generally small and cancel each other. This results in a method with low error compared to other anharmonic vibrational methods like second-order vibrational perturbation theory and to experiments. Several density functionals are tested using the method, and TPSS-D3, an inexpensive nonempirical density functional with dispersion corrections, is found to perform surprisingly well. Diffuse basis functions are found to be important for the accuracy of computed frequencies. Finally, the computation of vibrational properties like isotropic transition polarizabilities and the universality of the localized normal mode for terminal alkynes are demonstrated.

Thermal analysis of nanofluid magnetic flow on a rotating disk in the presence of radiation considering response surface method

This study examines Titania nanofluids that conduct electricity and are combined with various base fluids, focusing on various aspects. A magnetic field analyzes a continuous flow of nanofluid over a rotating disk that is positioned at an angle with a three-dimensional geometry assumption. The continuity, momentum, and energy balance equations are established, transformed using similarity variables, and solved through Akbari–Ganji Method (AGM) and Homotopy Perturbation Method (HPM) semi-analytical methods. The effects of several parameters, such as Magnetic parameter (M), Hall parameter (m), Porosity parameter ([Formula: see text]), Radiation parameter (Rd), and Thickness of liquid ([Formula: see text]), are examined through a graphical representation of state variables, including skin friction and Nusselt number. Additionally, the Response Surface Method (RSM) method has been utilized to simultaneously display the effects of porosity parameters and various factors across the disk. The findings showed that the magnetic, porosity, Hall, and thickness parameters had a significant impact on the behavior of the nanofluid. Specifically, the magnetic field had a strong influence on the velocity and temperature distribution of the fluid as it flowed over the rotating disk, while changes in porosity, Hall, and thickness parameters also affected these state variables.

Visuo-frontal interactions during social learning in freely moving macaques

Social interactions represent a ubiquitous aspect of our everyday life that we acquire by interpreting and responding to visual cues from conspecifics1. However, despite the general acceptance of this view, how visual information is used to guide the decision to cooperate is unknown. Here, we wirelessly recorded the spiking activity of populations of neurons in the visual and prefrontal cortex in conjunction with wireless recordings of oculomotor events while freely moving macaques engaged in social cooperation. As animals learned to cooperate, visual and executive areas refined the representation of social variables, such as the conspecific or reward, by distributing socially relevant information among neurons in each area. Decoding population activity showed that viewing social cues influences the decision to cooperate. Learning social events increased coordinated spiking between visual and prefrontal cortical neurons, which was associated with improved accuracy of neural populations to encode social cues and the decision to cooperate. These results indicate that the visual-frontal cortical network prioritizes relevant sensory information to facilitate learning social interactions while freely moving macaques interact in a naturalistic environment.

Variational representations of annealing paths: Bregman information under monotonic embedding

Variational representations of annealing paths: Bregman information under monotonic embedding

Time Variability Correction of CMIP6 Climate Change Projections

AbstractAccurate projections of climate change and associated extreme events under differing emission scenarios are linked to realistic representations of the temporal variability of the atmosphere at a variety of time scales, for example, annual, seasonal, synoptic, and daily. Here a new method is employed to explicitly quantify a model's ability to accurately represent covariance at and between differing time scales. From our global‐scale analysis, on average, raw Coupled Model Intercomparison Project Phase 6 (CMIP6) models misrepresent temporal variances at differing time scales for maximum temperature (tasmax) by a considerable margin, particularly at 183‐, 92‐ and 46‐day time scales. To ameliorate such variability errors, we propose a novel Time Variability Correction (TVC) method that corrects temporal covariances while preserving the essential time‐event sequence of the model simulations. We adopt a model‐as‐truth framework to evaluate the effectiveness of the TVC method under future forcing conditions when applied to daily tasmax simulations from 23 CMIP6 models for 1% of the global grid cells. TVC‐corrected temperatures generally show improved matching of temporal variance and lag correlations in the out‐of‐sample projection period compared to simple mean‐corrected projections. By imparting more realistic temporal‐correlations to model series, TVC is expected to improve the projections of extreme events associated with persistent heat, such as heatwaves. Applying TVC to future temperature projections using actual observations significantly increases the temperature variance in most middle to high latitude land regions in the Northern Hemisphere, while decreasing it in most low to middle latitude land regions, compared to simple mean‐corrected projections.

Application of a Simple Diffusivity Formulation to Examine Regime Transition and Jet–Eddy Energy Partitioning in Quasi-Geostrophic Turbulence

Abstract This study uses a simple diffusivity formulation to examine flow regime transition and jet–eddy energy partitioning in two-layer quasigeostrophic turbulence. Guided by simulations, the formulation is empirically constructed so that the diffusivity is bounded by a f-plane asymptote (Df) in the limit of vanishing β (termed drag-controlled) while reduced to a drag-independent scaling (Dβ) of Lapeyre and Held toward large β (termed β-controlled). Good agreement is found for diffusivities diagnosed from simulations with both quadratic and linear drag and in 2D turbulence. From the formulation, a regime diagram is readily constructed, with Df/Dβ = 1 separating the drag-controlled and β-controlled regimes. The diagram also sets the parameter range where an eddy velocity scaling is applicable. The quantitative representations of eddy variables then enable a reasonably skillful theory for zonal jet speed to be developed from energy balance. It is shown that, using Df/Dβ ≥ 10, a state where eddy statistics are approximately drag insensitive could be identified and interpreted using wave-damping competitions in slowing an inverse cascade. However, contrary to an existing hypothesis, the energy dissipation in such a state is not dominated by zonal jets. A modest revision for a way to maintain balance while keeping eddies drag insensitive is proposed. In the regime diagram, a subspace of zonostrophic condition, defined as jet dissipation surpassing eddy, is further quantified. It is demonstrated that a rough scaling could help interpret how the relative importance of jet and eddy dissipation varies across the parameter space.

Enhancing hydrologic modelling through the representation of traditional rainwater harvesting systems: A case study of water tanks in South India

AbstractWater tanks as traditional rainwater harvesting systems for agriculture are widely distributed in South India. They have a strong impact on hydrological processes, affecting streamflow in rivers as well as evapotranspiration. This study aims at an accurate representation of water harvesting systems in a hydrologic model to improve model performance and assessment of the catchment water balance. To this end, spatio‐temporal variations of water bodies between the years 2016 and 2018 and the months of January and May 2017 were derived from Sentinel‐2 satellite data to parameterize the water tanks (reservoir) parameters in the Soil and Water Assessment Tool (SWAT+) model of the Adyar basin, Chennai, India. Approximately 16% of the basin is covered by water tanks. The initial model performance was evaluated for two model setups, with and without water tanks. The best model run was selected with a multi‐metric approach comparing observed and modelled monthly streamflow for 5000 model runs. The final model evaluation was carried out by comparing estimated water body areas by the model and remote sensing observations for January to May 2017. The results showed that representing water tanks in the hydrologic model led to an improvement in the representation of the seasonal variations of streamflow for the whole simulation period (2004–2018). The model performance was classified as good and very good for the calibration (2004–2011) and validation (2012–2018) periods as NSE varies between 0.67 and 0.85, KGE varies between 0.65 and 0.72, PBIAS varies between −24.1 and −23.6, and RSR varies between 0.57 and 0.39. The best fit was shown for the high and middle flow segments of the hydrograph where the coefficient of determination (R2) ranges from 0.81 to 0.97 and 0.75 to 0.81, respectively. The monthly variation of water body areas in 2017 estimated by the hydrologic model was consistent with changes observed in remote sensing surveys. In summary, the water tank parametrization using remote sensing techniques enhanced the hydrologic model's efficiency and applicability for future studies.

Resource constrained neural network training

Modern applications of neural-network-based AI solutions tend to move from datacenter backends to low-power edge devices. Environmental, computational, and power constraints are inevitable consequences of such a shift. Limiting the bit count of neural network parameters proved to be a valid technique for speeding up and increasing efficiency of the inference process. Hence, it is understandable that a similar approach is gaining momentum in the field of neural network training. In the face of growing complexity of neural network architectures, reducing resources required for preparation of new models would not only improve cost efficiency but also enable a variety of new AI applications on modern personal devices. In this work, we present a deep refinement of neural network parameters limitation with the use of the asymmetric exponent method. In addition to the previous research, we study new techniques of floating-point variables limitation, representation, and rounding. Moreover, by leveraging exponent offset, we present floating-point precision adjustments without an increase in variables’ bit count. The proposed method allowed us to train LeNet, AlexNet and ResNet-18 convolutional neural networks with a custom 8-bit floating-point representation achieving minimal or no results degradation in comparison to baseline 32-bit floating-point variables.