Standard Feedforward Neural Network Research Articles

Hydrodynamic inline force prediction on vertical cylinders: a comparative study of neural network and its adaptive wavelets (wavenets)

The design of the neural network model and its adaptive wavelets (wavelet networks and wavenets) was used to estimate the wave-induced hydrodynamic inline force acting on a vertical cylinder. The data used to calibrate and validate the models were obtained from an experiment. In the brain, wavelet neural networks (WNNs) use wavelets to activate their hidden layers of neurons. In WNNs, both the position and dilation of the wavelets are optimized along with the weights. In one special approach to this kind of network construction, the position and dilation of the wavelets are fixed and only the weights of the network are optimized. In the present study, the neural network procedure and the above mentioned approach were employed to design a WNN, a so-called wavenet, using feed-forward neural network topology and its training method. Then, a comparison of these two methods was made. Numerical results demonstrate that both networks are capable of predicting hydrodynamic inline force. Furthermore, the combination of the neural network concept and the wavelet theory i.e. wavenet provides a more robust tool rather than standard feed-forward neural network, considering its more appropriate ability to predict any other data which the network had not experienced before. The results of this study can contribute to reducing the errors in future efforts to predict hydrodynamic inline force using WNNs, and thus improve the reliability of that prediction in comparison to the ANN and other methods. Therefore, this method can be applied to relevant engineering projects with satisfactory results.

ANN-QSAR Model for Virtual Screening of Androstenedione C-Skeleton Containing Phytomolecules and Analogues for Cytotoxic Activity Against Human Breast Cancer Cell Line MCF-7

The present study deals with the development of an artificial neural network based quantitative structure activity relationship (QSAR) model for virtual screening of active compounds which contain androstenedione carbonskeleton or their similar skeleton at the core. An empirical data modeling (with fitted data mapping) has been performed on the basis of bioassay record for human breast cancer cell line MCF7. The whole experimental data set was considered as test set. Standard feed-forward back-propagation neural network technique was applied to build the model. Leave-One- Out (LOO) cross-validation was performed to evaluate the performance of the model. The mapped model became the basis for selection best mapped compounds followed by development of Pharmacophore specific secondary QSAR model. In the present study, two best mapped molecules '4beta-hydroxy Withanolide-E' and '7, 8-Dehydrocalotropin' were used for development of the secondary QSAR model. These secondary-QSAR models were resulted with R2 LOOCV value 0.9845 and 0.9666 respectively. Docking studies, in silico phamacokinetic and toxicity analysis was also done for selected compounds. The screened compounds CID_73621, CID_16757497, CID_301751, CID_390666 and CID_46830222 were found with promising binding affinity value with aromatase with reference to the co-crystallized control compound androstenedione. Due to excellent extent of variance coverage in ANN based QSAR map model, it can be used as a robust non-linear QSAR model for androstenedione carbon-skeleton containing molecules and the protocol can be used to derive secondary QSAR models for other compounds set.

Groundwater modeling using hybrid of artificial neural network with genetic algorithm

Accurate estimates of groundwater level have a valuable effect in improving decision support systems of groundwater resources exploitation. The present study investigates the ability of a hybrid model of artificial neural network (ANN) and genetic algorithm (GA) in forecasting groundwater level in an individual well (target well). A standard feed forward networks (FFN) and recurrent neural networks (RNN) are utilized for performing the prediction task. Moreover, GA is used in order to determine the optimal structure of ANN (that is, number of neurons for each hidden layer). Air temperature, rainfall depth and groundwater levels in neighboring wells in Kerman plain (Kerman, Iran) were used as input data of the hybrid model. This study indicates that the ANN-GA model can be used successfully to forecast groundwater levels of individual wells. In addition, a comparative study of both hybrid models indicates that the feed forward networks performed better than the recurrent neural networks. Key words: Artificial neural network, feed forward networks, recurrent neural networks, genetic algorithm, groundwater level.

Hybrid genetic algorithm-neural network: Feature extraction for unpreprocessed microarray data

Suitable techniques for microarray analysis have been widely researched, particularly for the study of marker genes expressed to a specific type of cancer. Most of the machine learning methods that have been applied to significant gene selection focus on the classification ability rather than the selection ability of the method. These methods also require the microarray data to be preprocessed before analysis takes place. The objective of this study is to develop a hybrid genetic algorithm-neural network (GANN) model that emphasises feature selection and can operate on unpreprocessed microarray data. The GANN is a hybrid model where the fitness value of the genetic algorithm (GA) is based upon the number of samples correctly labelled by a standard feedforward artificial neural network (ANN). The model is evaluated by using two benchmark microarray datasets with different array platforms and differing number of classes (a 2-class oligonucleotide microarray data for acute leukaemia and a 4-class complementary DNA (cDNA) microarray dataset for SRBCTs (small round blue cell tumours)). The underlying concept of the GANN algorithm is to select highly informative genes by co-evolving both the GA fitness function and the ANN weights at the same time. The novel GANN selected approximately 50% of the same genes as the original studies. This may indicate that these common genes are more biologically significant than other genes in the datasets. The remaining 50% of the significant genes identified were used to build predictive models and for both datasets, the models based on the set of genes extracted by the GANN method produced more accurate results. The results also suggest that the GANN method not only can detect genes that are exclusively associated with a single cancer type but can also explore the genes that are differentially expressed in multiple cancer types. The results show that the GANN model has successfully extracted statistically significant genes from the unpreprocessed microarray data as well as extracting known biologically significant genes. We also show that assessing the biological significance of genes based on classification accuracy may be misleading and though the GANN's set of extra genes prove to be more statistically significant than those selected by other methods, a biological assessment of these genes is highly recommended to confirm their functionality.

Robust artificial intelligence tool for automatic start-up of the supplementary medium feeding in recombinant E. coli cultivations

One of the most important events in fed-batch fermentations is the definition of the moment to start the feeding. This paper presents a methodology for a rational selection of the architecture of an artificial intelligence (AI) system, based on a neural network committee (NNC), which identifies the end of the batch phase. The AI system was successfully used during high cell density cultivations of recombinant Escherichia coli. The AI algorithm was validated for different systems, expressing three antigens to be used in human and animal vaccines: fragments of surface proteins of Streptococcus pneumoniae (PspA), clades 1 and 3, and of Erysipelothrix rhusiopathiae (SpaA). Standard feed-forward neural networks (NNs), with a single hidden layer, were the basis for the NNC. The NN architecture with best performance had the following inputs: stirrer speed, inlet air, and oxygen flow rates, carbon dioxide evolution rate, and CO2 molar fraction in the exhaust gas.

A linear ridgelet network approach for fault diagnosis of analog circuit

A linear ridgelet network combining ridgelet, linear term and the standard feed-forward neural network used for diagnosing the faults of analog circuit is constructed, and a training algorithm based on the steepest gradient descent method and momentum method for this network and the procedure for diagnosing these faults are proposed. The resulting linear ridgelet network can learn more rapidly from training samples and handle more effectively the complicated fault information of circuit under test than wavelet network and ridgelet network, classifying these faults efficiently and correctly and achieving a high classification accuracy. The simulation results demonstrate the effectiveness of this approach.

Option Pricing With Modular Neural Networks

This paper investigates a nonparametric modular neural network (MNN) model to price the S&P-500 European call options. The modules are based on time to maturity and moneyness of the options. The option price function of interest is homogeneous of degree one with respect to the underlying index price and the strike price. When compared to an array of parametric and nonparametric models, the MNN method consistently exerts superior out-of-sample pricing performance. We conclude that modularity improves the generalization properties of standard feedforward neural network option pricing models (with and without the homogeneity hint).

Improving groundwater level forecasting with a feedforward neural network and linearly regressed projected precipitation

A module that uses neural networks was developed for forecasting the groundwater changes in an aquifer. A modified standard Feedforward Neural Network (FNN), trained with the Levenberg–Marquardt (LM) algorithm with five input variables (precipitation, temperature, runoff, groundwater level and specific yield) with a deterministic component, is used. The deterministic component links precipitation with the seasonal recharge of the aquifer and projects the seasonal average precipitations. A new algorithm is applied to forecast the groundwater level changes in Messara Valley, Crete, Greece, where groundwater level has been steadily decreasing due to overexploitation during the last 20 years. Results from the new algorithm show that the introduction of specific yield improved the groundwater level forecasting marginally but the linearly projected precipitation component drastically increased the window of forecasting up to 30 months, equivalent to five biannual time-steps.

An ANN‐based model for spatiotemporal groundwater level forecasting

AbstractThis paper evaluates the feasibility of using an artificial neural network (ANN) methodology for estimating the groundwater levels in some piezometers placed in an aquifer in north‐western Iran. This aquifer is multilayer and has a high groundwater level in urban areas. Spatiotemporal groundwater level simulation in a multilayer aquifer is regarded as difficult in hydrogeology due to the complexity of the different aquifer materials. In the present research the performance of different neural networks for groundwater level forecasting is examined in order to identify an optimal ANN architecture that can simulate the piezometers water levels. Six different types of network architectures and training algorithms are investigated and compared in terms of model prediction efficiency and accuracy. The results of different experiments show that accurate predictions can be achieved with a standard feedforward neural network trained usung the Levenberg–Marquardt algorithm. The structure and spatial regressions of the ANN parameters (weights and biases) are then used for spatiotemporal model presentation. The efficiency of the spatio‐temporal ANN (STANN) model is compared with two hybrid neural‐geostatistics (NG) and multivariate time series‐geostatistics (TSG) models. It is found in this study that the ANNs provide the most accurate predictions in comparison with the other models. Based on the nonlinear intrinsic ANN approach, the developed STANN model gives acceptable results for the Tabriz multilayer aquifer. Copyright © 2008 John Wiley & Sons, Ltd.

NEURAL NETWORK–BASED CLASSIFICATION OF NORMAL AND ABNORMAL PULMONARY FUNCTION USING SPIROMETRIC MEASUREMENTS

In this work, the classification of pulmonary function into normal and abnormal conditions is attempted using neural networks and spirometric measurements. The pulmonary function data (N = 229) for this study were obtained from volunteers using commercially available spirometers by adopting standard data acquisition protocol and recording conditions. The data were then subjected to neural network–based training (N = 159) and analysis (N = 70). The classification was carried out using standard feed-forward neural network and backpropagation algorithm. The outputs were then validated through sensitivity and specificity measurements together with clinical observation. The results confirmed the effectiveness of the neural network–based classification of spirometric data into normal and abnormal conditions. The sensitivity and specificity were found to be 89.25% and 82.25%, respectively. Furthermore, it seems that this method is useful in assessing the pulmonary function dynamics in cases with incomplete data and data with poor recordings. In this paper, the methodology, data collection procedure, and neural network–based analysis and results are described in detail.

Option Pricing With Modular Neural Networks

This paper applies a non-parametric modular neural network (MNN) model to price the S&P-500 European call options. The modules are based on time to maturity and moneyness of the options. The option price function of interest is homogenous of degree one with respect to the underlying index price and the strike price. We find that modularity improves the generalization properties of standard feedforward neural network option pricing models (with and without the homogeneity hint), relative to the Black-Scholes model.

Prediction of continuous B‐cell epitopes in an antigen using recurrent neural network

B-cell epitopes play a vital role in the development of peptide vaccines, in diagnosis of diseases, and also for allergy research. Experimental methods used for characterizing epitopes are time consuming and demand large resources. The availability of epitope prediction method(s) can rapidly aid experimenters in simplifying this problem. The standard feed-forward (FNN) and recurrent neural network (RNN) have been used in this study for predicting B-cell epitopes in an antigenic sequence. The networks have been trained and tested on a clean data set, which consists of 700 non-redundant B-cell epitopes obtained from Bcipep database and equal number of non-epitopes obtained randomly from Swiss-Prot database. The networks have been trained and tested at different input window length and hidden units. Maximum accuracy has been obtained using recurrent neural network (Jordan network) with a single hidden layer of 35 hidden units for window length of 16. The final network yields an overall prediction accuracy of 65.93% when tested by fivefold cross-validation. The corresponding sensitivity, specificity, and positive prediction values are 67.14, 64.71, and 65.61%, respectively. It has been observed that RNN (JE) was more successful than FNN in the prediction of B-cell epitopes. The length of the peptide is also important in the prediction of B-cell epitopes from antigenic sequences. The webserver ABCpred is freely available at www.imtech.res.in/raghava/abcpred/.

Evaluation of Neural Network Architectures for MPEG-4 Video Traffic Prediction

Multimedia applications and particularly MPEG-coded video streams are becoming major traffic components in high speed networks. Traffic prediction is important in enhancing the reliable operation over these networks. However, MPEG video traffic exhibits periodic correlation structure and a complex bit rate distribution, making prediction difficult. Neural networks can effectively be used to overcome such problem. In the literature, the problem has been mostly evaluated using standard feed-forward neural networks. However, a significant improvement can be expected using different types of neural networks. In this paper, six separate neural network predictors (including feed-forward) that can predict the basic frame types of MPEG-4: I, P, and B are developed and evaluated using long entertainment and broadcast video sequences. The performance is also compared to the widely used linear predictor. Comparison with results published in a recent work is also presented.

Special issue: Recent developments in hybrid intelligent systems

The human-centric facet of intelligent systems, which are inherently based upon the principle of hybridization of the individual information technologies, is definitely an interesting and highly promising avenue to follow. The diversity of communication channels between information systems and the environment is immense, and their effective realization does call for a truly heterogeneous nature of the underlying technologies and algorithmic frameworks. Consider, for instance, processing visual information, which became omnipresent in many applications including a broad range of inspection and manufacturing systems. Computer vision offered remarkable opportunities of processing therein. Fuzzy sets support a granular view at images and form a setting for their interpretation at the higher, more abstract level. Neural networks and evolutionary optimization provide badly needed learning abilities. The technologies mentioned so far, when combined into an inherently hybrid system, help address the needs of intelligent vision systems and build a coherent and versatile implementation. Managing knowledge at various levels, starting from the detailed pixel-level and moving to the far more general abstract level, becomes crucial to the success of such systems. This special issue is a testimony to the ongoing innovation and recent developments in hybrid information systems. The study by Shahram Jafari and Ray Jarvis focuses on implementing perception in the context of robotic eye-to-hand coordination. The authors show an integration of different novel concepts to perform scene analysis, hand-eye coordination and object manipulation for robotic tasks. The underlying architectural considerations rely on the usage of the neuro-fuzzy systems in the formation of scene analysis and object recognition. Next, Ghosh and Petkov, in their study, deal with the cognitive evaluation of contour based shape descriptors in image processing. In the sequel, first-order logical networks discussed by Kijsirikul and Lerdlamnaochai offer an interesting insight into the problems of logic-based neurocomputing. The proposed method, called First-Order Logical Neural Network (FOLNN), leads to hybrid structures that combine standard feed-forward neural networks and mechanisms of inductive learning. The study by Sehgal, Gondal, and Dooley is focused on an important and timely issue of handling multiple missing values of genetic expression in problems of analysis and classification of microarray data. The authors offer a mechanism of a k-ranked covariance-based missing value imputation and demonstrate its superiority over the k-nearest neighbor method.

Are Artificial Neural Networks White Boxes?

In this paper, we introduce a novel Mamdani-type fuzzy model, referred to as the all-permutations fuzzy rule base (APFRB), and show that it is mathematically equivalent to a standard feedforward neural network. We describe several applications of this equivalence between a neural network and our fuzzy rule base (FRB), including knowledge extraction from and knowledge insertion into neural networks.

Automated classification of eddy current signatures during manual inspection

This article presents a novel method for automatic evaluation of flaws during a manual eddy current (EC) inspection procedure. In manual scanning, a signal related to impedance change in the complex plane is affected by large and unpredictable variations of scanning speed and alterations of probe position. This paper introduces a robust EC signal normalization method using non-linear filtration based on evaluation of the distance between consecutive samples in the complex plane and median based tracking of the EC signature. Feature extraction was performed using normalized Fourier and complex discrete wavelet descriptors. The classification was performed using six different methods: nearest-mean classifier, k-nearest neighborhood classifier, the standard multi-layer feed forward neural network with backpropagation, radial basis network, support vector machines and the adaptive-network-based fuzzy inference system. The method was tested on a single frequency instrument with an absolute probe and a dual frequency instrument with a probe for fast testing of rivets in layered structures. The results obtained using this approach demonstrate the effectiveness of the proposed system.

Groundwater level forecasting using artificial neural networks

A proper design of the architecture of Artificial Neural Network (ANN) models can provide a robust tool in water resources modeling and forecasting. The performance of different neural networks in a groundwater level forecasting is examined in order to identify an optimal ANN architecture that can simulate the decreasing trend of the groundwater level and provide acceptable predictions up to 18 months ahead. Messara Valley in Crete (Greece) was chosen as the study area as its groundwater resources have being overexploited during the last fifteen years and the groundwater level has been decreasing steadily. Seven different types of network architectures and training algorithms are investigated and compared in terms of model prediction efficiency and accuracy. The different experiment results show that accurate predictions can be achieved with a standard feedforward neural network trained with the Levenberg–Marquardt algorithm providing the best results for up to 18 months forecasts.

A Note on the Universal Approximation Capability of Support Vector Machines

The approximation capability of support vector machines (SVMs) is investigated. We show the universal approximation capability of SVMs with various kernels, including Gaussian, several dot product, or polynomial kernels, based on the universal approximation capability of their standard feedforward neural network counterparts. Moreover, it is shown that an SVM with polynomial kernel of degree p − 1 which is trained on a training set of size p can approximate the p training points up to any accuracy.

Protein secondary structure: category assignment and predictability

In the last decade, the prediction of protein secondary structure has been optimized using essentially one and the same assignment scheme known as DSSP. We present here a different scheme, which is more predictable. This scheme predicts directly the hydrogen bonds, which stabilize the secondary structures. Single sequence prediction of the new three category assignment gives an overall prediction improvement of 3.1% and 5.1% compared to the DSSP assignment and schemes where the helix category consists of α-helix and 3 10-helix, respectively. These results were achieved using a standard feed-forward neural network with one hidden layer on a data set identical to the one used in earlier work.

An artificial neural network as a model for prediction of survival in trauma patients: validation for a regional trauma area.

To develop and validate an artificial neural network (ANN) for predicting survival of trauma patients based on standard prehospital variables, emergency room admission variables, and Injury Severity Score (ISS) using data derived from a regional area trauma system, and to compare this model with known trauma scoring systems. The study was composed of 10,609 patients admitted to 24 hospitals comprising a seven-county suburban/rural trauma region adjacent to a major metropolitan area. The data was generated as part of the New York State trauma registry. Study period was from January 1993 through December 1996 (1993-1994: 5,168 patients; 1995: 2,768 patients; 1996: 2,673 patients). A standard feed-forward back-propagation neural network was developed using Glasgow Coma Scale, systolic blood pressure, heart rate, respiratory rate, temperature, hematocrit, age, sex, intubation status, ICD-9-CM Injury E-code, and ISS as input variables. The network had a single layer of hidden nodes. Initial network development of the model was performed on the 1993-1994 data. Subsequent models were generated using the 1993, 1994, and 1995 data. The model was tested first on the 1995 and then on the 1996 data. The ANN model was tested against Trauma and Injury Severity Score (TRISS) and ISS using the receiver operator characteristic (ROC) area under the curve [ROC-A(z)], Lemeshow-Hosmer C-statistic, and calibration curves. The ANN showed good clustering of the data, with good separation of nonsurvivors and survivors. The ROCA(z) was 0.912 for the ANN, 0.895 for TRISS, and 0.766 for ISS. The ANN exceeded TRISS with respect to calibration (Lemeshow-Hosmer C-statistic: 7.4 for ANN; 17.1 for TRISS). The prediction of survivors was good for both models. The ANN exceeded TRISS in nonsurvivor prediction. An ANN developed for trauma patients using prehospital, emergency room admission data, and ISS gave good prediction of survival. It was accurate and had excellent calibration. This study expands our previous results developed at a single Level I trauma center and shows that an ANN model for predicting trauma deaths can be applied across hospitals with good results