Reduced Prediction Accuracy Research Articles

Digital mapping of soil biological properties and wheat yield using remotely sensed, soil chemical data and machine learning approaches

Mapping of soil properties by using novel machine learning (ML) algorithms and optimized environmental covariates is of great importance for agricultural management to enhance crop production. This research aimed at evaluating ML algorithms to predict spatial distribution of soil biological properties and wheat yield in the Southwest of Iran. Topsoil samples (0–30 cm) were collected from a total of 60 sampling locations and wheat grain yield (plot 1 × 1 m) was recorded at each location. Soil properties including urease (Ur), alkaline phosphatase (AP), basal respiration (BR), microbial biomass carbon (MBC), soil organic carbon (SOC), MBC:SOC ratio, and metabolic quotient (qCO2) were measured. At the first step, Random Forest (RF) model was employed to predict soil biological properties by using terrain attributes, remote sensing indices and soil properties as covariates. In this step, both Variance Inflation Factor (VIF) and Pearson regression were applied to select the most important covariates in predicting soil biological properties and to decrease the dimension of the input space with considering no reduction in prediction accuracy. Secondly, wheat grain yield was modeled using six ML algorithms; they were optimized and evaluated in Caret package with 10-fold cross validation. Results showed the highest prediction accuracy for qCO2 (R2adj = 0.80) and the lowest for BR (R2adj = 0.23). Compared to environmental predictors, soil covariates had a greater effect in modeling Ur, qCO2, MBC and MBC:SOC ratio, while, for AP and BR, bands 6 and Chanel Network Base Level were the most important factors, respectively. In prediction of wheat grain yield, both Stochastic Gradient Boosting (SGB) and RF models outperformed with R2adj of 0.89 and 0.88, respectively. Results indicated that the Ur and AP played the major roles in predicting wheat grain yield and explaining its spatial variability. Our modeling results suggested that soil biological properties and yield can be estimated easily with reasonable accuracy. Overall, their high resolution maps may be useful for decision makers, stakeholders and applicants in agricultural management practices towards precision agriculture.

Quantitative prediction accuracies derived from laser-induced breakdown spectra using optimized multivariate submodels

The accuracy of laser-induced breakdown spectroscopy (LIBS) methods for analyzing geological samples is improved when calibration standards and unknown targets are compositionally similar. A recent study suggests that customized submodels can be used to optimize calibration datasets to achieve more accurate predictions [1]. In practice, this is difficult to implement because the errors inherent in the methods used for sorting unknown targets by composition may affect how successfully this matching can occur. Moreover, creation of submodels intrinsically reduces the size of the dataset on which the model is trained, which has been shown to reduce prediction accuracy. This paper uses LIBS spectra of 2990 unique rock powder standards to compare the accuracy of 1) submodels generated for each element over its geochemical range, 2) submodels created using SiO2 content only, 3) submodels created using the ratio of Si(II)/Si(I) emission lines to group spectra by a proxy for approximate plasma temperature, and 4) models created using all data. Results indicate that prediction accuracies are not always improved by creating submodels because subdividing a dataset to optimize calibrations will always result in a smaller database available for each submodel, and the reduced training set size negatively affects accuracy. Customized LIBS standards for specific applications might overcome this problem in cases where the matrix is similar and the expected concentration range is known. But in a majority of geochemical applications, submodel approaches are only useful in improving prediction accuracies when the initial database is itself extensive enough to support large, robust submodel calibration suites.

Real-time milk analysis integrated with stacking ensemble learning as a tool for the daily prediction of cheese-making traits in Holstein cattle

Cheese-making traits in dairy cattle are important to the dairy industry but are difficult to measure at the individual level because there are limitations on collecting phenotypic information. Mid-infrared spectroscopy has its advantages, but it can only be used during monthly milk recordings. Recently, in-line devices for real-time analysis of milk quality have been developed. The AfiLab recording system (Afimilk) offers significant benefits as phenotypes can be collected from each cow at each milking session. The objective of this study was to assess the potential of integrating AfiLab real-time milk analyzer measures with the stacking ensemble learning technique using heterogeneous base learners for the in-line daily monitoring of cheese-making traits in Holstein cattle with a view to developing a precision livestock farming system for monitoring the technological quality of milk. Data and samples for wet-laboratory analyses were collected from 499 Holstein cows belonging to 2 farms where the AfiLab system was installed. The traits of concern were 9 milk coagulation traits [3 milk coagulation properties (MCP), and 6 curd firming traits (CFt)], and 7 cheese-making traits [3 cheese yield (CY) traits, and 4 milk nutrient recovery in the curd (REC) traits]. The near-infrared AfiLab spectral data and on-farm information (days in milk and parity) were used to assess the predictive ability of different statistical methods [elastic net (EN), gradient boosting machine (GBM), extreme gradient boosting (XGBoost), and artificial neural network (ANN)] across different cross-validation scenarios. These statistical methods were considered the base learners, which were then combined in a stacking ensemble learning. Results indicate that including information on the cows (days in milk and parity) in the AfiLab infrared prediction increased its accuracy by 10.3% for traditional MCP, 13.8% for curd firming, 9.8% for CY, and 11.2% for REC traits compared with those obtained from near-infrared AfiLab alone. The statistical approaches exhibited high prediction accuracies (R2) averaged across the cross-validation scenarios for traditional MCP (0.58 for ANN, 0.55 for EN and GBM, 0.52 for XGBoost, and 0.62 for stacking ensemble), CFt (0.55 for ANN, 0.54 for EN and GBM, 0.53 for XGBoost, and 0.61 for stacking ensemble), and similar R2 averages for CY and REC (0.55 for ANN, 0.54 for EN and GBM, 0.53 for XGBoost, and 0.61 for stacking ensemble). The ANN approach was more accurate than the other base learners (EN, GBM, and XGBoost) and improved accuracy across cross-validation scenarios on average by 7% for traditional MCP, 5% for CFt, 8% for CY, and 7% for REC. The stacking ensemble method improved prediction accuracy by 3% to 31% for traditional MCP, 2% to 26% for CFt, 1% to 38% for CY traits, and 2% to 27% for REC traits compared with the base learners. The prediction accuracies of the different approaches evaluated tended to decrease from the 10-fold cross-validation to the independent validation scenario, although there was a smaller reduction in prediction accuracy with the stacking ensemble learning technique across all the cross-validation scenarios. Our results show that combining in-line on-farm information with stacking ensemble machine learning represents an effective alternative for obtaining robust daily predictions of milk cheese-making traits.

Genomic prediction models for traits differing in heritability for soybean, rice, and maize

BackgroundGenomic selection is a powerful tool in plant breeding. By building a prediction model using a training set with markers and phenotypes, genomic estimated breeding values (GEBVs) can be used as predictions of breeding values in a target set with only genotype data. There is, however, limited information on how prediction accuracy of genomic prediction can be optimized. The objective of this study was to evaluate the performance of 11 genomic prediction models across species in terms of prediction accuracy for two traits with different heritabilities using several subsets of markers and training population proportions. Species studied were maize (Zea mays, L.), soybean (Glycine max, L.), and rice (Oryza sativa, L.), which vary in linkage disequilibrium (LD) decay rates and have contrasting genetic architectures.ResultsCorrelations between observed and predicted GEBVs were determined via cross validation for three training-to-testing proportions (90:10, 70:30, and 50:50). Maize, which has the shortest extent of LD, showed the highest prediction accuracy. Amongst all the models tested, Bayes B performed better than or equal to all other models for each trait in all the three crops. Traits with higher broad-sense and narrow-sense heritabilities were associated with higher prediction accuracy. When subsets of markers were selected based on LD, the accuracy was similar to that observed from the complete set of markers. However, prediction accuracies were significantly improved when using a subset of total markers that were significant at P ≤ 0.05 or P ≤ 0.10. As expected, exclusion of QTL-associated markers in the model reduced prediction accuracy. Prediction accuracy varied among different training population proportions.ConclusionsWe conclude that prediction accuracy for genomic selection can be improved by using the Bayes B model with a subset of significant markers and by selecting the training population based on narrow sense heritability.

PERFORMANCE ANALYSIS OF AN EFFICIENT FRAMEWORK FOR INTRUSION DETECTION SYSTEM USING DATA MINING TECHNIQUES

In the midst of the COVID-19 epidemic crisis, due to the tremendous development of mobile and internet technologies, the excessive growth in cyber-crime makes networksurity a major concern. As a result, individuals and companies are gradually moving towards the use of Intrusion Detection System (IDS), as it plays a persuasive role in monitoring and detecting the traffic of a network. However, high dimensional data affect the performance of IDS by reducing prediction accuracy, increasing false positive rate and classification time. Hence the focus of this research work is to develop a novel framework by integrating Auto – Bi Level (ABL) Classification with Double Filtering Fine Tuning – Ensemble Hybrid (DFFT-EH) feature selection. The experiments are conducted using NSL- KDD a benchmark intrusion detection dataset and it is proved that the proposed framework performs well with good accuracy, less false positive rate and less classification time when compared with voting ensemble classifier and other existing standard algorithms.

Abstract PR-13: Potential effect on racial/ethnic disparities of removing racial/ethnic variables from risk models: The example of lung-cancer screening

Abstract Background: Some uses of “race correction” in clinical algorithms and prediction models unfairly reduce access to care, resulting in calls to remove racial/ethnic variables from all models and algorithms. However, for models that are based on unbiased, high-quality, and plentiful data, removing racial/ethnic variables may reduce prediction accuracy for minorities. We model racial/ethnic disparities in screening eligibility from augmenting USPSTF-2021 guidelines (ages 50-80, ≥20 pack-years, ≤15 quit-years) to also include individuals selected by an NCCN-recommended risk model that includes race (PLCOM2012) versus the same model with race/ethnicity removed (PLCOM2012_NoRace). Methods: We used previously published methodology to model the performance of lung cancer screening using 6915 ever-smokers ages 50-80 from the US-representative 2015 National Health Interview Survey (NHIS). Individuals were considered eligible for screening if they are eligible by USPSTF-2021 guidelines or by PLCOM2012 (“USPSTF+PLCOM2012”), versus being eligible by USPSTF-2021 or PLCOM2012_NoRace (“USPSTF+PLCOM2012_NoRace”). Both models used the NCCN-recommended ≥1.3% 6-year risk-threshold for eligibility. We evaluated model accuracy (average percent over/under-estimation) by race/ethnicity, estimated the proportion of life-years gainable achieved by each eligible cohort (LYG), and evaluated the LYG disparity (difference in LYG between whites and each minority). Results: USPSTF+PLCOM2012 and USPSTF+PLCOM2012_NoRace identified similar numbers of minorities as eligible for screening (~2.7 million). However, USPSTF+PLCOM2012_NoRace selected 125% more Hispanic-Americans and 31% less African-Americans. LYG disparities decreased using USPSTF+PLCOM2012_NoRace versus USPSTF+PLCOM2012 for Hispanic Americans (LYG: 33% to 29%). However, LYG disparities for African Americans increased (LYG: 16% to 18%). PLCOM2012 underestimated lung cancer risk by 49% for Hispanic-Americans, whereas PLCOM2012_NoRace performed well (4% overestimation). However, PLCOM2012underestimated risk in African-Americans by only 6%, PLCOM2012_NoRace underestimated risk in African-Americans by 36%. Conclusion: The model that was most accurate for a minority group was projected to reduce disparities the most for that group. Removing race from the PLCOM2012 model substantially underestimated risk for African-Americans and may increase disparities. Inexplicably, PLCOM2012 substantially underestimated risk in Hispanic-Americans despite including race/ethnicity, which was alleviated by removing race/ethnicity. Great care must be taken when removing racial/ethnic variables from models, because this will assign minorities risk estimates that may be largely, or entirely, based on the majority population. Citation Format: Corey D. Young, Li C. Cheung, Christine D. Berg, Patricia Rivera, Hilary A. Robbins, Anil K. Chaturvedi, Hormuzd A. Katki, Rebecca Landy. Potential effect on racial/ethnic disparities of removing racial/ethnic variables from risk models: The example of lung-cancer screening [abstract]. In: Proceedings of the AACR Virtual Conference: 14th AACR Conference on the Science of Cancer Health Disparities in Racial/Ethnic Minorities and the Medically Underserved; 2021 Oct 6-8. Philadelphia (PA): AACR; Cancer Epidemiol Biomarkers Prev 2022;31(1 Suppl):Abstract nr PR-13.

Real-estate price prediction with deep neural network and principal component analysis

Abstract Despite the wide application of deep neural networks (DNN) models, their application over small-sized real-estate price prediction is limited due to the reduced prediction accuracy and the high-dimensionality of the dataset. This study motivates small-sized real-estate agencies to take DNN-driven decisions using the available local dataset. To improve the high-dimensionality of real-estate price datasets and thus enhance the price-prediction accuracy of a DNN model, this paper adopts principal component analysis (PCA). The PCA benefits in improving the prediction accuracy of a DNN model are threefold: dimensionality reduction, dataset transformation and localisation of influential price features. The results indicate that, through the PCA-DNN model, the transformed dataset achieves higher accuracy (90%–95%) and better generalisation ability compared with other benchmark price predictors. The spatial and building age proved to have the most impact in determining the overall real-estate price. The application of PCA not only reduces the high-dimensionality of the dataset but also enhances the quality of the encoded feature attributes. The model is beneficial in real-estate and construction applications, where the absence of medium and big datasets decreases the price-prediction accuracy.

Neural‐network transformation models for counting processes

AbstractWhile many survival models have been invented, the Cox model and the proportional odds model are among the most popular ones. Both models are special cases of the linear transformation model. The linear transformation model typically assumes a linear function on covariates, which may not reflect the complex relationship between covariates and survival outcomes. Nonlinear functional form can also be specified in the linear transformation model. Nonetheless, the underlying functional form is unknown and mis‐specifying it leads to biased estimates and reduced prediction accuracy of the model. To address this issue, we develop a neural‐network transformation model. Similar to neural networks, the neural‐network transformation model uses its hierarchical structure to learn complex features from simpler ones and is capable of approximating the underlying functional form of covariates. It also inherits advantages from the linear transformation model, making it applicable to both time‐to‐event analyses and recurrent event analyses. Simulations demonstrate that the neural‐network transformation model outperforms the linear transformation model in terms of estimation and prediction accuracy when the covariate effects are nonlinear. The advantage of the new model over the linear transformation model is also illustrated via two real applications.

Relative permeability measurement of coal microchannels using advanced microchip technology

The gas–water flow in coal cleats is an important factor affecting coalbed methane production. However, difficulties are often encountered in the prediction of gas production. Limited understanding of two-phase flow behaviour in microchannels reduces prediction accuracy, often leading to overestimation. In this study, based on micro-computed tomography, advanced microchip technology is used to predict the optimal coal cleat in the polydimethylsiloxane chip. Gas and liquid injection experiments are carried out on the microchannels through the syringe pumps, as the two-phase flow morphology in the system can obtain flow information that cannot be provided by conventional laboratory measurements. The flow capacities of single-phase gas, water, and wetting liquid in the microchannels are compared, alongside determining the flow laws under different wettability and flow rates. The results show that in the single-phase fluid injection experiment, gas is more sensitive to the microchannel size effect than the liquid phase, and hydrophilic liquids flow readily. The gas–water relative permeability value in this study is significantly smaller than that obtained from conventional experiments, based on the steady-state method. This is a result of the gas–water forming a discontinuous plug flow in the microchannel, which produces higher resistance and reduces flow capacity. The occurrence of discontinuous flow regimes challenges the laminar flow assumption adopted in the steady-state method and suggests that conventional laboratory measurements of relative permeability measurements may misrepresent the physics of fluid flows in microchannels. The results also reveal that a hydrophilic environment and a suitable flow rate can increase the two-phase relative permeability. The results of this study enhance our understanding of and approach to core flooding measurement, with the differences in the relative permeability curves indicating that data uncertainty associated with the steady-state method is worthy of further research.

Probabilistic spatiotemporal solar irradiation forecasting using deep ensembles convolutional shared weight long short-term memory network

High-quality and reliable solar radiation spatiotemporal probabilistic prediction is very significant for the planning and management of solar energy. In this paper, convolutional shared weight long short-term memory network is proposed to deal with solar radiation spatiotemporal prediction problems. The model can minimize the training time of model without significantly reducing prediction accuracy. Furthermore, a spatiotemporal probabilistic prediction model that combines convolutional shared weight long short-term memory network and deep ensemble method is proposed to deal with the solar radiation probabilistic prediction problem. The model obtains the uncertainty estimation of predictands by adjusting the network structure and optimizes the uncertainty by employing a proper scoring rule. In addition, a combination of ensembles are used in the model to improve the robustness of probabilistic prediction. The new spatiotemporal probabilistic prediction model is applied to predict solar radiation in a real area of the United States. Five state-of-the-art models and seven evaluation indicators are used for comparation. The comparation results show the proposed model is able to provide accurate point prediction, reasonable prediction interval and reliable probabilistic prediction results for a whole area.

Machine Learning-Based Code Auto-Completion Implementation for Firmware Developers

With the advent of artificial intelligence, the research paradigm in natural language processing has been transitioned from statistical methods to machine learning-based approaches. One application is to develop a deep learning-based language model that helps software engineers write code faster. Although there have already been many attempts to develop code auto-completion functionality from different research groups, a need to establish an in-house code has been identified for the following reasons: (1) a security-sensitive company (e.g., Samsung Electronics) may not want to utilize commercial tools given that there is a risk of leaked source codes and (2) commercial tools may not be applicable to the specific domain (e.g., SSD firmware development) especially if one needs to predict unique code patterns and style. This research proposes a hybrid approach that harnesses the synergy between machine learning techniques and advanced design methods aiming to develop a code auto-completion framework that helps firmware developers write code in a more efficient manner. The sensitivity analysis results show that the deterministic design results in reducing prediction accuracy as it generates output in some unexpected ways, while the probabilistic design provides a list of reasonable next code elements in which one could select it manually to increase prediction accuracy.

Reducing moisture effects on soil organic carbon content prediction in visible and near-infrared spectra with an external parameter othogonalization algorithm

Field spectroscopy and other efficient hyperspectral techniques have been widely used to measure soil properties, including soil organic carbon (SOC) content. However, reflectance measurements based on field spectroscopy are quite sensitive to uncontrolled variations in surface soil conditions, such as moisture content; hence, such variations lead to drastically reduced prediction accuracy. The goals of this work are to (i) explore the moisture effect on soil spectra with different SOC levels, (ii) evaluate the selection of optimal parameter for external parameter othogonalization (EPO) in reducing moisture effect, and (iii) improve SOC prediction accuracy for semi-arid soils with various moisture levels by combing the EPO with machine learning method. Soil samples were collected from grassland regions of Inner Mongolia in North China. Rewetting laboratory experiments were conducted to make samples moisturized at five levels. Visible and near-infrared spectra (350–2500 nm) of soil samples rewetted were observed using a hand-held SVC HR-1024 spectroradiometer. Our results show that moisture influences the correlation between SOC content and soil reflectance spectra and that moisture has a greater impact on the spectra of samples with low SOC. An EPO algorithm can quantitatively extract information of the affected spectra from the spectra of moist soil samples by an optimal singular value. A SOC model that effectively couples EPO with random forest (RF) outperforms partial least-square regression (PLSR)-based models. The EPO–RF model generates better results with R2 of 0.86 and root-mean squared error (RMSE) of 3.82 g kg−1, whereas a PLSR model gives R2 of 0.79 and RMSE of 4.68 g kg−1.

On the cross-population generalizability of gene expression prediction models.

The genetic control of gene expression is a core component of human physiology. For the past several years, transcriptome-wide association studies have leveraged large datasets of linked genotype and RNA sequencing information to create a powerful gene-based test of association that has been used in dozens of studies. While numerous discoveries have been made, the populations in the training data are overwhelmingly of European descent, and little is known about the generalizability of these models to other populations. Here, we test for cross-population generalizability of gene expression prediction models using a dataset of African American individuals with RNA-Seq data in whole blood. We find that the default models trained in large datasets such as GTEx and DGN fare poorly in African Americans, with a notable reduction in prediction accuracy when compared to European Americans. We replicate these limitations in cross-population generalizability using the five populations in the GEUVADIS dataset. Via realistic simulations of both populations and gene expression, we show that accurate cross-population generalizability of transcriptome prediction only arises when eQTL architecture is substantially shared across populations. In contrast, models with non-identical eQTLs showed patterns similar to real-world data. Therefore, generating RNA-Seq data in diverse populations is a critical step towards multi-ethnic utility of gene expression prediction.

Short-term runoff prediction with GRU and LSTM networks without requiring time step optimization during sample generation

Runoff forecasting is an important approach for flood mitigation. Many machine learning models have been proposed for runoff forecasting in recent years. To reconstruct the time series of runoff data into a standard machine learning dataset, a sliding window method is usually used to pre-process the data, with the size of the window as a variable parameter which is commonly referred to as the time step. Conventional machine learning methods, such as artificial neural network models (ANN), require optimization of the time step because both too small and too large time steps reduce prediction accuracy. In this work two popular variants of Recurrent Neural Network (RNN) named Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks were employed to develop new data-driven flood forecasting models. GRU and LSTM models are in theory able to filter redundant information automatically, and therefore a large time step is expected to not reduce prediction accuracy. The three models (LSTM, GRU, and ANN) were applied to simulate runoff in the Yutan station control catchment, Fujian Province, Southeast China, using hourly discharge measurements of one runoff station and hourly rainfall of four rainfall stations from 2000 to 2014. Results show that the prediction accuracy of LSTM and GRU models increases with increasing time step, and eventually stabilizes. This allows selection of a relatively large time step in practical runoff prediction without first evaluating and optimizing the time step required by conventional machine learning models. We also show that LSTM and GRU models perform better than ANN models when the time step is optimized. GRU models have fewer parameters and less complicated structures compared to LSTM models, and our results show that GRU models perform equally well as LSTM models. GRU may be the preferred method in short term runoff predictions since it requires less time for model training.

An integrated modelling approach examining the influence of goals, habit and learning on choice using visual attention data

Previous economics literature has explored the role of visual attention on choice in isolation without accounting for other influences such as habits and goals or learning effects, nor their interrelationship. In this paper, we: (i) develop a novel joint framework to explore the relationship between visual attention, observed heterogeneity from stated habits and goals, and choice outcomes while accounting for shorter- and longer-term learning effects; and (ii) investigate whether accounting for these relationships improves model predictive power and behavioral insights. The empirical analysis used an eye-tracked discrete choice experiment on sugar-sweetened beverage purchasing (n = 152 adults with 20 choice tasks). Results suggest that habits, goals, and shorter-term learning are key drivers of information acquisition whereas cumulative choices (longer-term learning) affect subsequent choice outcome. Importantly, ignoring the joint relationship between habits, visual attention and choice may exaggerate the role of visual attention, leading to incorrect behavioral insights and reduced prediction accuracy.

Prediction of shape distortions during forming and welding of a double-curved strip geometry in alloy 718

The finite element method (FEM) has considerably contributed to the development of advanced manufacturing methods for metal structures. The prediction of the final shape of a component is of great interest to the manufacturing industry. The level of demand may increase due to multistage processes. Therefore, including all steps of the manufacturing chain in the simulations is a key to being successful. This has been done for a long time in the stamping industry, which involves sequences of forming, trimming, and springback. However, more complex manufacturing procedures that include assembling of formed parts with forgings and castings via welding have been modelled with simplifications, resulting in a reduced prediction accuracy. In the present study, a double-curved part manufactured from alloy 718 is formed at 20 °C and laser-welded using the bead-on-plate procedure. The coupling of different manufacturing analyses, including cold forming, trimming, result mapping, welding, cooling, and springback, is achieved using LS-DYNA. Additionally, the effect of adding a damage and failure model in the forming simulation is studied. The results of the forming analysis are used as inputs for the material model *MAT_CWM in the welding simulation. The anisotropic thermomechanical properties of alloy 718 are determined at temperatures up to 1000 °C. Encouraging agreement is found between the model predictions and the results of forming and welding tests. The findings underscore the importance of including the material history and accurate process conditions along the manufacturing chain to both the prediction accuracy of shape distortions, and to the potential of the industry.

Neural Networks on an FPGA and Hardware-Friendly Activation Functions

This paper describes our implementation of several neural networks built on a field programmable gate array (FPGA) and used to recognize a handwritten digit dataset—the Modified National Institute of Standards and Technology (MNIST) database. We also propose a novel hardware-friendly activation function called the dynamic Rectifid Linear Unit (ReLU)—D-ReLU function that achieves higher performance than traditional activation functions at no cost to accuracy. We built a 2-layer online training multilayer perceptron (MLP) neural network on an FPGA with varying data width. Reducing the data width from 8 to 4 bits only reduces prediction accuracy by 11%, but the FPGA area decreases by 41%. Compared to networks that use the sigmoid functions, our proposed D-ReLU function uses 24% - 41% less area with no loss to prediction accuracy. Further reducing the data width of the 3-layer networks from 8 to 4 bits, the prediction accuracies only decrease by 3% - 5%, with area being reduced by 9% - 28%. Moreover, FPGA solutions have 29 times faster execution time, even despite running at a 60× lower clock rate. Thus, FPGA implementations of neural networks offer a high-performance, low power alternative to traditional software methods, and our novel D-ReLU activation function offers additional improvements to performance and power saving.

A Parallel Multilevel Feature Selection algorithm for improved cancer classification

Biological data is prone to grow exponentially, which consumes more resources, time and manpower. Parallelization of algorithms could reduce overall execution time. There are two main challenges in parallelizing computational methods. (1) Biological data is multi-dimensional in nature. (2). Parallel algorithms reduce execution time, but with the penalty of reduced prediction accuracy. This research paper targets these two issues and proposes the following approaches. (1) Vertical partitioning of data along feature space and horizontal partitioning along samples in order to ease the task of data parallelism. (2) Parallel Multilevel Feature Selection (M-FS) algorithm to select optimal and important features for improved classification of cancer sub-types. The selected features are evaluated using parallel Random Forest on Spark, compared with previously reported results and also with the results of sequential execution of same algorithms. The proposed parallel M-FS algorithm was compared with existing parallel feature selection algorithms in terms of accuracy and execution time. The results reveal that parallel multilevel feature selection algorithm improved cancer classification resulting into prediction accuracy ranging from ∼85% to ∼99% with very high speed up in terms of seconds. On the other hand, existing sequential algorithms yielded prediction accuracy of ∼65% to ∼99% with execution time of more than 24 hours.

Irrigation, crop stress and drainage reduction under uncertainty: A scenario study

Two thirds of human water use is linked to agricultural practices including crop irrigation. Furthermore, excess irrigation leads to drainage problems. For this reason, reduced irrigation strategies need to be implemented to protect water resources. However, low irrigation may lead to crop water stress. A fast and inexpensive way to predict the necessary amount of irrigation required is by a model-based approach. With this approach, it is possible to explore the relation between irrigation, crop water stress and drainage. However, parameter uncertainty can reduce prediction accuracy. Therefore, the aims of this research were: (1) to develop and test a methodology that allows the analysis of uncertainty sources in irrigation strategies (2) to identify how much irrigation can be reduced while maintaining a low risk of crop stress, and (3) to explore the influence of uncertainty in soil parameters and evapotranspiration on model predictions. Results from a realistic case considered in this study indicated that, while maintaining a low risk of crop stress (<1 %), it is possible to reduce drainage (by 88 %) and water use (22 %) for a conventional irrigation strategy. This reduction is dependent on the type of risk aversion strategy and is specific for a case scenario where variations are certain.

The hybrid protein interactome contributes to rice heterosis as epistatic effects.

Heterosis is the phenomenon in which hybrid progeny exhibits superior traits in comparison with those of their parents. Genomic variations between the two parental genomes may generate epistasis interactions, which is one of the genetic hypotheses explaining heterosis. We postulate that protein-protein interactions specific to F1 hybrids (F1 -specific PPIs) may occur when two parental genomes combine, as the proteome of each parent may supply novel interacting partners. To test our assumption, an inter-subspecies hybrid interactome was simulated by in silico PPI prediction between rice japonica (cultivar Nipponbare) and indica (cultivar 9311). Four-thousand, six-hundred and twelve F1 -specific PPIs accounting for 20.5% of total PPIs in the hybrid interactome were found. Genes participating in F1 -specific PPIs tend to encode metabolic enzymes and are generally localized in genomic regions harboring metabolic gene clusters. To test the genetic effect of F1 -specific PPIs in heterosis, genomic selection analysis was performed for trait prediction with additive, dominant and epistatic effects separately considered in the model. We found that the removal of single nucleotide polymorphisms associated with F1 -specific PPIs reduced prediction accuracy when epistatic effects were considered in the model, but no significant changes were observed when additive or dominant effects were considered. In summary, genomic divergence widely dispersed between japonica and indica rice may generate F1 -specific PPIs, part of which may accumulatively contribute to heterosis according to our computational analysis. These candidate F1 -specific PPIs, especially for those involved in metabolic biosynthesis pathways, are worthy of experimental validation when large-scale protein interactome datasets are generated in hybrid rice in the future.