Search

Overdriving technique enlarges the desired change of the pixel value, the error in general compression methods is enlarged at the same time. Therefore, we propose a novel Adaptive Quantization Coding (AQC) to reduce the error in compression for overdriving technique reducing motion blur. Considering hardware implementation, we develop a hybrid image coding which uses color transform first, and then uses AQC to compress luminance data as well as Block Truncation Coding (BTC) to compress chrominance data. The simulation results shown that the average PSNR was improved 5.676dB as compared with the result of BTC, and the average SD of error was reduced 50.2% than that in the BTC. The proposed algorithm is implemented with the verilog HDL and synthesized with the synopsys design compiler using 0.13μm Samsung Library. [ABSTRACT FROM AUTHOR]

For the service based Grid applications, the applications may be integrated by using the Grid services across Internet, thus we should balance the load for the applications to enhance the resource's utility and increase the throughput. To overcome the problem, one effective way is to make use of load balancing. Kinds of load balancing middleware have already been applied successfully in distributed computing. However, they don't take the services types into consideration and for different services requested by clients the workload would be different out of sight. Furthermore, traditional load balancing middleware uses the fixed and static replica management and uses the load migration to relieve overload. Therefore, we put forward an autonomic replica management infrastructure to support fast response, hot-spot control and balanced resource allocation among services. Corresponding simulation tests are implemented and results indicated that this model and its supplementary mechanisms are suitable to service based Grid applications. [ABSTRACT FROM AUTHOR]

By using middleware, we can satisfy the urgent demands of performance, scalability and availability in current distributed service-oriented applications. However to the complex applications, the load peak may make the system suffer extremely high load and the response time may be decreased for this kind of fluctuate. Therefore, to utilize the services effectively especially when the workloads fluctuate frequently, we should make the system react to the load fluctuate gradually and predictably. Many existing load balancing middleware use the dampening technology to make the load to be predicative. However, distributed systems are inherently difficult to manage and the dampening factor cannot be treated as static and fixed. The dampening factor should be adjusted dynamically according to different load fluctuate. So we have proposed a new technique based on machine learning for adaptive and flexible load prediction mechanism based on our load balancing middleware. [ABSTRACT FROM AUTHOR]

Image classification is a challenging problem in organizing a large image database. However, an effective method for such an objective is still under investigation. This paper presents a method based on wavelet for image classification with adaptive processing of data structures. After decomposed by wavelet, the features of an image can be reflected by the wavelet coefficients. Therefore, the nodes of tree representation of images are represented by distribution of histograms of wavelet coefficient projections. 2940 images derived from seven original categories are used in experiments. Half of the images are used for training neural network and the other images used for testing. The classification rate of training set is 90%, and the classification rate of testing set is 87%. The promising results prove the proposed method is very effective and reliable for image classification. [ABSTRACT FROM AUTHOR]

Artificial Neural Networks (ANNs) which are derived from Biological Neural Networks (BNNs) are enhanced by many advanced mathematical techniques and have become powerful tools for solving complicated engineering problems. Integrating BNNs with mature ANNs is a very effective method of solving intricate biological problems and explaining neurophysiological data. In this paper we propose a neural network model that explains how the brain processes visual information about impending collisions with an object - in particular, how time-to-collision information is caculated in the brain. The model performs extremely well as a result of incorporating physiological data with the methods involved in the development of ANNs. By implementing this novel compuational neural network model, the results of the simulation demonstrate that this integrative approach is a very useful and efficient way to deal with complicated problems in neural computation. [ABSTRACT FROM AUTHOR]

Optimizing a plasma impedance match process requires construction of prediction model. In this study, generalized regression neural network (GRNN) combined with genetic algorithm (GA) was used to build a match prediction model. A real-time match monitor system was used to collect steady match positions according to a statistical experimental design. GA-GRNN models were compared to GRNN and statistical regression models. Compared to GRNN models, GA-GRNN models demonstrated improved predictions of about 81% and 77% for the impedance and phase positions, respectively. With respect to statistical regression models, GA-GRNN models yielded an improvement of about 80% and 78%, respectively. Moreover, for either model type, the improvements for the training errors were more than about 90% for both positions. [ABSTRACT FROM AUTHOR]

Gene regulatory inference from time series gene expression data, generated from DNA microarray, has become increasingly important in investigating genes functions and unveiling fundamental cellular processes. Computational methods in machine learning and neural networks play an active role in analyzing the obtained data. Here, we investigate the performance of particle swarm optimization (PSO) on the reconstruction of gene networks, which is modeled with recurrent neural networks (RNN). The experimental results on a synthetic data set are presented to show the parameter effects of PSO on RNN training and the effectiveness of the proposed method in revealing the gene relations. [ABSTRACT FROM AUTHOR]

We developed a high-speed concurrent support vector machine (CSVM) processor for real-time nonlinear channel equalization. All phases of the recognition process, including kernel computing, learning, and recall of the support vector machine (SVM) are performed on a single chip. The concurrent operation of this CSVM using a parallel architecture of elements allows it to achieve high speed. The hardware-friendly kernel adatron (KA) SVM learning algorithms are embedded on a chip. The results of the nonlinear channel equalization obtained by the KA algorithm are compared with those obtained by the quadratic programming (QP) method. The CSVM using the KA learning algorithm is designed and implemented using the FPGA chip. The CSVM processor performs 20% faster than the existing SVM processors. [ABSTRACT FROM AUTHOR]

An all-optical neural network based on coupled ring lasers is proposed in this paper. Each laser in the network has a different wavelength, representing one neuron. The network status is determined by the wavelength of the network's light output. Inputs to the network are in the optical power domain. The nonlinear threshold function required for neural-network operation is achieved optically by interaction between the lasers. A simple laser model developed in the paper has illuminated the behavior of the coupled lasers. An experimental system is implemented using single mode fiber optic components at wavelengths near 1550 nm. A number of functions are implemented to demonstrate the practicality of the new network. From the experiment, a conclusion can be obtained that the neural network is particularly robust against input wavelength variations. [ABSTRACT FROM AUTHOR]

Research and development efforts are ongoing in the implementation of a totally integrated fiber-is-the sensor for structural and environmental sensing using the highly efficient Brillouin effect. Sensitivity has further improved by incorporating the smart structures for elevated actuation power, and logic memory or neural network capabilities. The presence of Brillouin and Kerr nonlinearity in the fiber together with inherent feedback delay has been shown to lead to instabilities and ultimately optical chaos. The deterministic nature of this process and its multiple bifurcations may be suppressed by a number of scheme, such as continues interference feedback, while its extremely large memory capacity exploited. These memories can be estimated as an optical logic function used for all-optic in-line switching, channel selection, amplification, oscillation in optical communication, optical logic elements in optical computation based on neural networks applications and hardware implementation. [ABSTRACT FROM AUTHOR]

We propose a silicon synapse for spiking neural network application. In this endeavor, two major issues are addressed: the structure of the synapse and the associated behavior. This synaptic structure is basically a charge transfer device comprising of two Metal-Oxide-Semiconductor (MOS) capacitors the first of which stores the weight and the second controls its reading. In this work, simulation results prove that the proposed synapse captures the intrinsic dynamics of the biological synapse and exhibits a spike characteristic. The device operates at very low power and offers the potential for scaling to massively parallel third generation hardware neural networks. [ABSTRACT FROM AUTHOR]

Neural Networks are usually implemented in software on sequential machines but when implemented in hardware, they are extremely fast due to the massive parallelism inherent in the hardware devices. Implementation of Neural Networks in Programmable Logic Devices such as FPGAs (Field Programmable Gate Arrays) gives us more flexibility since these devices are reconfigurable and their design can be altered whenever needed. The design proposed in this paper shows the implementation of perceptron neural network in FPGAs for the character recognition problem. The characters here are the English language alphabets which are input to the network and after training; they are tested for recognition. Each alphabet is tested for three different fonts. After implementation, the simulations are done and performance issues of the design are analyzed.The post-layout simulation gives excellent results even if some noise is introduced to the input patterns. [ABSTRACT FROM AUTHOR]

Many hardware implementations cannot execute the software MLPs' applications using weight of floating-point data, because hardware design of MLPs usually uses fixed-point arithmetic for high speed and small area. The hardware design using fixed-point arithmetic has two important drawbacks which are low accuracy and flexibility. Therefore, we propose a fully-pipelining architecture of MLPs using floating-point arithmetic in order to solve these two problems. Thus our design method can implement the MLPs having the processing speed improved by optimizing the number of hidden nodes in a repeated processing. We apply a software application of MLPs-based text detection that is computed to be 1722120 times for text detection of a 1152×1546 sized image to hardware implementation. Our preliminary result shows a performance enhancement of about eleven times faster using a fully-pipelining architecture than the software application. [ABSTRACT FROM AUTHOR]

In this paper, hybrid neural network processor (HANNP) is designed in VLSI. The HANNP has RISC based architecture leading to an effective general digital signal processing and artificial neural networks computation. The architecture of a HANNP including the general digital processing units such as 64-bit floating-point arithmetic unit (FPU), a control unit (CU) and neural network processing units such as artificial neural computing unit (NNPU), specialized neural data bus and interface unit, etc. The HANNP is modeled in Veilog HDL and implemented with FPGA. Character recognition problems and Kohonen self-organization problems are applied to the proposed HANNP to justify its applicability to real engineering problems. [ABSTRACT FROM AUTHOR]

In this paper we present the results of neural network hardware in-the-loop training for an analogue Local Cluster Neural Network (LCNN) chip. We use a Probabilistic Random Weight Change (PRWC) algorithm that is a combination of the random weight change and simulated annealing algorithms. We applied the PRWC algorithm to in-the-loop training for multi-dimensional function approximations and for predictions. We discuss the training strategy and the experiment results. [ABSTRACT FROM AUTHOR]

This paper proposes a reconfigurable architecture for VLSI implementation of BP neural networks with on-chip learning. Basing on systolic arrays, this architecture can flexibly adapt to neural networks with different scales, transfer functions or learning algorithms by reconfiguration of basic processing components,. Three kinds of reconfigurable processing units (RPU) are proposed firstly basing on the analysis of neural network's reconfiguration. Secondly, the paper proposes a reconfigurable systolic architecture and the method of mapping BP networks into this architecture. The implementation of an instance on FPGA is introduced in the last. The results show that this flexible architecture can also achieve a high learning speed of 432M CUPS (Connection Updated Per Second) at 100MHz using 22 multipliers. [ABSTRACT FROM AUTHOR]

This paper proposes an efficient hardware architecture for a function generator suitable for an artificial neural network (ANN). A spline-based approximation function is designed that provides a good trade-off between accuracy and silicon area, whilst also being inherently scalable and adaptable for numerous activation functions. This has been achieved by using a minimax polynomial and through optimal placement of the approximating polynomials based on the results of a genetic algorithm. The approximation error of the proposed method compares favourably to all related research in this field. Efficient hardware multiplication circuitry is used in the implementation, which reduces the area overhead and increases the throughput. [ABSTRACT FROM AUTHOR]

Neural network is employed to construct the power macromodel of complementary metal-oxide-semiconductor (CMOS) integrated circuits. In contrast to previous modeling approaches, it does not require empirically constructed specialized analytical equations for the power macromodel, and obtained statistics of a circuit's primary outputs simultaneously. It is suitable for power estimation in core-based systems-on-chips (SoCs) with pre-designed blocks. In experiments with the ISCAS-85 circuits, the average absolute relative error of the macromodel was below 5.0% for not only the average power dissipation, but also the maximum power dissipation. [ABSTRACT FROM AUTHOR]

Application of feedforward neural networks for integrated circuit (IC) modeling is presented. In order to accurately describe IC behaviors, a set of improved equations for dynamic feedforward neural networks has been utilized for IC modeling. The rationality of the improved equations is elucidated by analyzing the relation between the circuits and the equation parameters. Through some special choices of the neuron nonlinearity function, the feed- forward neural networks can themselves be represented by equivalent circuits, which enables the direct use of neural models in existing analogue circuit simulators. Feedforward neural network models for some static and dynamic systems are obtained and compared. Simulated results are included to illustrate the accuracy of the neural networks in circuit modeling. [ABSTRACT FROM AUTHOR]

The principal components of near infrared spectroscopy were calculated by nonlinear iterative partial least squares (NIPALS). The best number of principal components was determined by cross-validation method. Thus, limited principal components that free from noise and orthogonal each other were obtained. After standardization, these principal components were used as input nodes of back propagation artificial neural networks (B-P ANN). ANN was used to build nonlinear model. In the method, the data of whole spectra can be fully utilized, the best principal components free from noise and nonlinear model obtained, the iterative time of B-P ANN shorted strongly, and better calibration model can be obtained. The method has been applied to quantitatively determine the starch of barely. The calibration and prediction correlation coefficients are 0.982 and 0.945; the relative standard deviations are 1.81% and 2.80%, respectively. [ABSTRACT FROM AUTHOR]

A filtering method aided by neural network to improve the maneuvering target tracking performance is proposed in this paper. Based on unscented Kalman filter, the unscented particle filter (UPF) has more accurate proposal distribution and better approximation to non-linear tracking problem than other Sequential Monte-Carlo methods. The neural network is constructed and trained by the maneuvering features, and the outputs of NN are used as acceleration control parameters to correct model parameters. Simulation results show the performance of UPF aided by NN is much improved than extensive Kalman filter. [ABSTRACT FROM AUTHOR]

Combining the time-frequency location and multiple-scale analysis of wavelet transform with the nonlinear mapping and self-learning of neural networks, an error decomposition method in dynamic measuring system is proposed. According to whole-system dynamic accuracy theory, the whole-error model of dynamic measuring system is given, and then the whole-error is decomposed into sub-errors by wavelet neural networks, which are traced so that units of the system are found which generate these errors and error transmission characteristic is controlled. [ABSTRACT FROM AUTHOR]

This paper presents the application of artificial neural network (ANN) in saturator. Phase Doppler Anemometry (PDA) is utilized to investigate the distribution of water droplets diameter and velocity in the saturator. The data obtained from experiment is used as input-output of ANN. Before using ANN method, some prerequisites have to be processed, including the selection of the number of input and output variables, hidden layer neurons, the network architecture and the normalization of data etc. The results indicate that the trained ANN can provide accurate prediction values which agree with real experimental data closely. [ABSTRACT FROM AUTHOR]

Expertise acquisition is always the obstacle and bottleneck in the development of intelligent design system. In order to generalize and accumulate the expertise and experience of simulation analysis and experiments, the intelligent design system of sugarcane harvester is introduced. In the intelligent system of sugarcane harvester, the neural network is applied to overcome the difficulty of knowledge acquisition (KA). In this study, the application of neural network in the system is illustrated, including data predisposal, generation and management of the knowledge. An example is given to explain the application as well. The research shows using neural network can simplify the procedure of knowledge acquisition. It can also evaluate and forecast the performance of sugarcane harvester in design phrase. And it is beneficial to enhance the development success rate of the digital product and to lessen the development cost of physical prototype. [ABSTRACT FROM AUTHOR]

The biological wastewater treatment plant, which uses microbial community to remove organic matter and nutrients in wastewater, is known as its nonlinear behavior and uncertainty to operate. In spite of strong needs of automatic monitoring of nutrients, it is thought that tremendous expense may be required to install equipments related with remote control system, especially on-line sensors for monitoring organic and nutrient concentrations in the treatment processes. In this research, as a cost-effective tool for replacing expensive on-line sensor, PNN(Polynomial Neural Network) models were developed to estimate the NOx-N and ammonia concentrations by only using on-line values of ORP, DO and pH at the wastewater treatment plant. Developed PNN model could estimate the NOx-N and ammonia profile well. However, the error was increased at the first anoxic period of the first sub-cycle and NOx-N accumulation was occurred at the sub-cycle. To deal with those errors, the rule-base-compensator was developed based on operational knowledge. [ABSTRACT FROM AUTHOR]

An application based on BP neural network is conducted on evaluating the state of metal loss in gas conducting pipelines. We employ network to quantify the defects of metal loss in outer pipe spool according to the wm, lm and hm of MFL (magnetic flux leakage), which are the width, length and depth of the area of abnormal magnetic data. After quantifying, we obtain w, l and d, which are deemed to the width, length and depth of the defect. The data to train our networks are developed from MFL data which are picked by Pipe Pig moving in service pipe with a constant velocity parallel to pipe axis. [ABSTRACT FROM AUTHOR]

The models of dual modality densitometry were developed. It can be used for measuring the gas volume fraction and water volume fraction in oil water gas pipe flow. The models are complex. In order to solve models, it often uses simplified models. This reduces measurement precision. The method of measuring gas and water volume fraction using neural networks was presented. The simulation data was gotten using Geant4. The radial basis function networks were trained and tested on computer simulation data. The results show that networks predicted gas volume fraction fit true gas fraction well and water volume fraction has some deviations. [ABSTRACT FROM AUTHOR]

A new method for arbitrary 3d-object reconstruction in unknown environment is proposed in this paper. The implicit surface is reconstructed based on radial basis functions network from range scattered data. For the property of locality of radial basis function, the method is fast and robust with respect to large data. Furthermore, an adapted K-Means algorithm is used to reduce RBF centers for reconstruction. Experiment results show that the presented approach is helpful in speed improvement and is a good solution for large data reconstruction. [ABSTRACT FROM AUTHOR]

The issue of pre-splitting blasting design is investigated in this paper. According to the principle of expert system and neural networks, Visual C++ 6.0 and SQL Sever 2000 are employed to develop hybrid expert system of pre-splitting blasting design based on BP neural networks. The proposed expert system is a coupling system of engineering database and three-layered BP neural networks, which can be applied into cutting excavation of the expressway. The experiments show that the system can enhance the reliability of in-site pre-splitting blasting design scheme, efficiency and quality of construction. [ABSTRACT FROM AUTHOR]

Earthquakes will do great harms to the people, to estimate the future earthquake situation in Chinese mainland is still an open issue. There have been previous attempts to solve this problem by using artificial neural networks. In this paper, a novel algorithm named MIFEB is proposed to improve the estimation accuracy by combing bagging of neural networks with mutual information based feature selection for its individuals. MIFEB is compared with the general case of bagging on UCI data sets, then, MIFEB is used to forecast the seismicity of strong earthquakes in Chinese mainland, computation results show that MIFEB obtains higher accuracy than other several methods like bagging of neural networks and single neural networks do. [ABSTRACT FROM AUTHOR]

With the accumulating of the strong earthquakes records, it becomes practicable to achieve the more accurate attenuation relationships. Based on the seismic records of West American, the Radial Basis Function (RBF) and Back Propagation (BP) artificial neural networks model are respectively constructed for three-dimensional seismic parameters attenuation relationship. The RBF model is nice fitting for the training data, although it has great errors on other tested points. While the BP model is not good than the RBF model for the training data, it possesses a better consecutive property in the whole area. It is a proper neural network model for the problem. After training with the selected records, the Neural Networks (NN) shows a good fitting with the training records. And it is easy to construct three-dimensional model to predict the attenuation relationship. In order to demonstrate the efficiency of the presented methodology, the contrast is discussed for the results of the BP model and three typical traditional attenuation formulae. [ABSTRACT FROM AUTHOR]

Electronic noses (E-nose) have gained popularity in various applications such as food inspection, cosmetics quality control [1], toxic vapor detection to counter terrorism, detection of Improvised Explosive Devices (IED), narcotics detection, etc. In the paper, we summarized our results on the application of Support Vector Machines (SVM) to gas detection and classification using E-nose. First, based on experimental data from Jet Propulsion Lab. (JPL), we created three different data sets based on different pre-processing techniques. Second, we used SVM to detect gas sample data from non-gas background data, and used three sensor selection methods to improve the detection rate. We were able to achieve 85% correct detection of gases. Third, SVM gas classifier was developed to classify 15 different single gases and mixtures. Different sensor selection methods were applied and FSS & BSS feature selection method yielded the best performance. [ABSTRACT FROM AUTHOR]

We proposed a fast cloud detection approach for the geophysical data from Moderate Resolution Imaging Spectroradiometer (MODIS), a premium instrument aboard on NASA's satellite Terra to study clouds and aerosols. Previous pixel-based classifiers have been developed for remote-sensing instruments using various machine learning techniques, such as artificial neural networks (ANNs), support vector machines (SVMs). However, their computational costs are very expensive. Our novel approach integrated image segmentation and SVMs together to achieve the similar classification accuracy while using much less computation costs. It exploited the homogeneous property in local spatial sub-regions and used radiance information from sub-regions, rather than pixels, to build classifiers. The experimental results showed the proposed approach not only greatly speed up the classification training procedure, but also provide insights for domain experts to reveal different cloud types. [ABSTRACT FROM AUTHOR]

In this paper a novel optimization approach is presented. Network architecture and connection weights of neural networks (NN) are evolved by a particle swarm optimization (PSO) method, and then the appropriate network architecture and connection weights are fed into back-propagation (BP) networks. The ensemble strategy is carried out by simple averaging. The applied example is built with monthly mean rainfall of the whole area in Guangxi, China. The results show that the proposed approach can effectively improves convergence speed and generalization ability of NN. [ABSTRACT FROM AUTHOR]

Lattice data have two different scale spatial properties: global depen- dence property and local fluctuation property. For lattice data space-time autoregressive modeling, to reduce influence of spatial fluctuation on prediction accuracy of neural networks, all regions are partitioned into several subareas by an improved k-means algorithm based on spatial contiguity relation. Some partition criteria are proposed to evaluate different partition schemes and the optimal scheme has the least spatial fluctuation and significant spatial dependent within each subarea. Each multi-layer perceptrons (MLPs) network is modeled respectively for each subarea, and the output nodes are the prediction values at time t of an attribute for all regions in a subarea, and the input nodes are observations before time t of this subarea itself and neighboring regions. As a case study, all local models are tested and compared with a single global MLPs network by one-step-ahead predicting of an epidemic dataset, and the results indicate that local NN model has better prediction performance than the global NN model. [ABSTRACT FROM AUTHOR]

A FNN prediction model based on hedonic price theory to estimate the appropriate price level for a new real estate is proposed. The model includes a database storing hedonic characteristics and coefficients affecting the real estate price level from recently sold projects that are representative in the local environment. The experimental result shows that the fuzzy neural network prediction model has strong function approximation ability and is suitable for real estate price prediction depending on the quality of the available data. [ABSTRACT FROM AUTHOR]