In this paper, we proposed the novel method for the recognition of English calling cards by using the contour tracking algorithm and the enhanced fuzzy RBF (Radial Basis Function) neural networks. The recognition of calling cards consists of the extraction phase of character areas and the recognition phase of extracted characters. In the extraction phase, first of all, noises are removed from the images of calling cards, and the feature areas including character strings are separated from the calling card images by using the horizontal smearing method and the 8-directional contour tracking method. And using the image projection method the feature areas are split into the areas of individual characters. We also proposed the enhanced fuzzy RBF neural network that organizes the middle layer effectively by using the enhanced fuzzy ART neural network adjusting the vigilance parameter dynamically according to the similarity between patterns. In the recognition phase, the proposed fuzzy neural network was applied to recognize individual characters. Our experiment result showed that the proposed recognition algorithm has higher success rate of recognition and faster learning time than the conventional RBF network based recognitions.

Applying the formerly proposed classification framework for first order line search optimization techniques we introduce novel superlinear first order line search methods. Novelty of the methods lies in the line search subproblem. The presented line search subproblem features automatic step length and momentum adjustments at every iteration of the algorithms realizable in a single step calculation. This keeps the computational complexity of the algorithms linear and does not harm the stability and convergence of the methods. The algorithms have none or linear memory requirements and are shown to be convergent and capable of reaching the superlinear convergence rates. They were practically applied to artificial neural network training and compared to the relevant training methods within the same class. The simulation results show satisfactory performance of the introduced algorithms over the standard and previously proposed methods.

Evaluation of cyclic transitions in the discrete-time neural networks with antisymmetric and circular interconnection weights has been derived in an asymptotic mathematical form. The type and the number of limit cycles generated by circular networks, in which each neuron is connected only to its nearest neurons, have been investigated through analytical method. The results show that the estimated numbers of state vectors generating n- or 2n-periodic limit cycles are an exponential function of (1.6)n for a large number of neuron, n. The sufficient conditions for state vectors to generate limit cycles of period n or 2n are also given.

This paper proposes a novel feature selection method for the fuzzy neural networks and presents an application example for 'personalized' facial expression recognition. The proposed method is shown to result in a superior performance than many existing approaches.

In this paper, we present a predictive control method, based on Fuzzy Neural Network (FNN), for the control of chaotic systems without precise mathematical models. In our design method, the parameters of both predictor and controller are tuned by a simple gradient descent scheme, and the weight parameters of the FNN are determined adaptively throughout system operations. In order to design the predictive controller effectively, we describe the computing procedure for each of the two important parameters. In addition, we introduce a projection matrix for determining the control input, which decreases the control performance function very rapidly. Finally, we depict various computer simulations on two representative chaotic systems (the Duffing and Hénon systems) so as to demonstrate the effectiveness of the new chaos control method.

A problem of estimating biopotential sources in the brain based on EEG signals observed on the scalp is known as an important inverse problem of electrophysiology. Usually there is no closed-form solution for this problem and it requires iterative techniques such as the Levenberg-Marquardt algorithm. Considering the nonlinear properties of inverse problem, and signal to noise ratio inherent in EEG signals, a back propagation neural network has been recently proposed as a solution. In this paper, we investigated the properties of neural networks and its localization accuracy for single dipole source localization. Based on the results of extensive studies, we concluded the neural networks are highly feasible in single-source localization with a small number of electrodes (18 electrodes), also examined the usefulness of this method for clinical application with a case of epilepsy.

A new multilayer artificial neural network learning algorithm based on the pattern search method is proposed. The learning algorithm is designed to provide a very simple and effective means of searching the minima of an objective function directly without any knowledge of its derivatives. We test this algorithm on benchmark problems, such as exclusive-or (XOR), parity and alphabetic character learning problems. For all problems, the systems are shown to be trained efficiently by our algorithm. As a simple direct search algorithm, it can be applied to hardware implementations easily.

We propose a minimal model of neuronal burst-firing that can be considered as a modification and extention of the Bonhoeffer-van der Pol (BVP) model. By using linear stability analysis we show that one of the equilibrium points of the fast subsystem is a saddle point which divides the phase plane into two regions. In one region all phase trajectories approach a limit cycle and in the other they approach a stable equilibrium point. The slow subsystem describes a slowly varying inward current. Various types of bursting phenomena are presented by using bifurcation analysis. The simplicity of the model and the variety of firing modes are the biggest advantages of our model with obvious applications in understanding underlying mechanisms of generation of neuronal firings and modeling oscillatory neural networks.

The wavelet transform (WT) has recently emerged as a powerful tool for image compression. In this paper, a new image compression technique combining the genetic algorithm (GA) and grey-based competitive learning network (GCLN) in the wavelet transform domain is proposed. In the GCLN, the grey theory is applied to a two-layer modified competitive learning network in order to generate optimal solution for VQ. In accordance with the degree of similarity measure between training vectors and codevectors, the grey relational analysis is used to measure the relationship degree among them. The GA is used in an attempt to optimize a specified objective function related to vector quantizer design. The physical processes of competition, selection and reproduction operating in populations are adopted in combination with GCLN to produce a superior genetic grey-based competitive learning network (GGCLN) for codebook design in image compression. The experimental results show that a promising codebook can be obtained using the proposed GGCLN and GGCLN with wavelet decomposition.

In this paper, we propose the CMOS implementation of neuron models for an artificial auditory neural network. We show that when voltage is added directly to the control terminal of the basic circuit of the hardware neuron model, a change in the output firing is observed. Next, based on this circuit, a circuit that changes with time is added to the control terminal of the basic circuit of the hardware neuron model. As a result, a neuron model is constructed with ON firing, adaptation firing, and repetitive firing using CMOS. Furthermore, an improved circuit of a neuron model with OFF firing using CMOS which has been improved from the previous model is also constructed.

The jaw movements can be measured by estimating the position and orientation of two small permanent magnets attached on the upper and lower jaws. It is a difficult problem to estimate the positions and orientations of the magnets from magnetic field because it is a typical inverse problem. The back propagation neural networks (BPNN) are applicable to solve this problem in short processing time. But its precision is not enough to apply to practical measurement. In the other hand, precise estimation is possible by using the nonlinear least-square (NLS) method. However, it takes long processing time for iterative calculation, and the solutions may be trapped in the local minima. In this paper, we propose a precise and fast measurement system which makes use of the estimation algorithm combining BPNN with NLS method. In this method, the BPNN performs an approximate estimation of magnet parameters in short processing time, and its result is used as the initial value of iterative calculation of NLS method. The cost function is solved by Gauss-Newton iteration algorithm. Precision, processing time and noise immunity were examined by computer simulations. These results shows the proposed system has satisfactory ability to be applied to practical measurement.

This paper presents a flexible learning algorithm for the binary neural network that can realize a desired Boolean function. The algorithm determines hidden layer parameters using a genetic algorithm. It can reduce the number of hidden neurons and can suppress parameters dispersion. These advantages are verified by basic numerical experiments.

This paper investigates the behavior of one-dimensional discrete-time binary cellular neural networks with both the A- and B-templates and gives the necessary and sufficient conditions for the above network to be stable for unspecified fixed boundaries.

In this paper we treat a novel adaptation strength according to neighborhood ranking of self-organizing neural networks with the objective of avoiding the initial dependency of reference vectors, which is related to the strength in the neural-gas network suggested by Martinetz et al. The present approach exhibits the effectiveness in the average distortion compared to the conventional technique through numerical experiments. Furthermore the present approach is applied to image data and the validity in employing as an image coding system is examined.

A new algorithm for synthesizing binary threshold neural networks (BTNNs) is proposed. A binary (Boolean) input-output mapping that can be represented by minimal sum-of-product (MSP) terms is initially obtained from training data. The BTNN is then synthesized based on an MSP term grouping method. As a result, a fast and optimal realization of a BTNN can be obtained. Examples of both feedforward and recurrent BTNN synthesis used in a parallel processing architecture are given and compared with other existing methods.

Associative memories composed of sparsely interconnected neural networks (SINNs) are suitable for analog hardware implementation. However, the sparsely interconnected structure also gives rise to a decrease in the capability of SINNs for associative memories. Although this problem can be solved by increasing the number of interconnections, the hardware cost goes up rapidly. Therefore, we propose associative memories consisting of multilayer perceptrons (MLPs) with 3-valued weights and SINNs. It is expected that such MLPs can be realized at a lower cost than increasing interconnections in SINNs and can give each neuron in SINNs the global information of an input pattern to improve the storage capacity. Finally, it is confirmed by simulations that our proposed associative memories have good performance.

A method of learning for multi-layer artificial neural networks is proposed. The learning model is designed to provide an effective means of escape from the Backpropagation local minima. The system is shown to escape from the Backpropagation local minima and be of much faster convergence than simulated annealing techniques by simulations on the exclusive-or problem and the Arabic numerals recognition problem.

We propose hysteresis neural network solving combinatorial optimization problems, Box Puzzling Problem. Hysteresis neural network searches solutions of the problem with nonlinear dynamics. The output vector becomes stable only when it corresponds with a solution. This system does never become stable without satisfying constraints of the problem. After estimating hardware calculating time, we obtain that numerical calculating time increases extremely comparing with hardware time as problem's scale increases. However the system has possibility of limit cycle. Though it is very hard to remove limit cycle completely, we propose some methods to remove this phenomenon.

This paper describes the implementation of a proposed image filter into a Discrete-Time Cellular Neural Network (DT-CNN). The three stages that compose the filter are described, showing that the resultant filter is capable of (1) erasing or detecting several concentric shapes simultaneously, (2) thresholding and (3) thinning of gray-scale images. Because the DT-CNN has to fill certain conditions for this filter to be implemented, it becomes a modified version of a DT-CNN. Those conditions are described and also experimental results are clearly shown.

Multilayer feedforward neural network (MFNN) trained by the backpropagation (BP) algorithm is one of the most significant models in artificial neural networks. MFNNs have been used in many areas of signal and image processing due to high applicability. Although they have been implemented as analog, mixed analog-digital and fully digital VLSI circuits, it is still difficult to realize their hardware implementation with the BP learning function efficiently. This paper describes a special BP algorithm for the logic oriented neural network (LOGO-NN) which we have proposed as a sort of MFNN with quantized weights and multilevel threshold neurons. Both weights and neuron outputs are quantized to integer values in LOGO-NNs. Furthermore, the proposed BP algorithm can reduce high precise calculations. Therefore, it is expected that LOGO-NNs with BP learning can be more effectively implemented as digital type circuits than the common MFNNs with the classical BP. Finally, it is shown by simulations that the proposed BP algorithm for LOGO-NNs has good performance in terms of the convergence rate, convergence speed and generalization capability.