This paper proposes a prediction method for non-stationary time series data with time varying parameters. A modular structured type neural network is newly introduced for the purpose of grasping the changing property of time varying parameters. This modular structured neural network is constructed by the hierarchical combination of each neural network (NNT: Neural Network for Prediction of Time Series Data) and a neural network (NNW: Neural Network for Prediction of Weights). Next, we propose a reasonable method for determination of the length of the local stationary section by using the additive learning ability of neural networks. Finally, the validity and effectiveness of the proposed method are confirmed through simulation and actual experiments.

This paper compares signal classification performance of multilayer neural networks (MLNNs) and linear filters (LFs). The MLNNs are useful for arbitrary waveform signal classification. On the other hand, LFS are useful for the signals, which are specified with frequency components. In this paper, both methods are compared based on frequency selective performance. The signals to be classified contain several frequency components. Furthermore, effects of the number of the signal samples are investigated. In this case, the frequency information may be lost to some extent. This makes the classification problems difficult. From practical viewpoint, computational complexity is also limited to the same level in both methods.IIR and FIR filters are compared. FIR filters with a direct form can save computations, which is independent of the filter order. IIR filters, on the other hand, cannot provide good signal classification deu to their phase distortion, and require a large amount of computations due to their recursive structure. When the number of the input samples is strictly limited, the signal vectors are widely distributed in the multi-dimensional signal space. In this case, signal classification by the LF method cannot provide a good performance. Because, they are designed to extract the frequency components. On the other hand, the MLNN method can form class regions in the signal vector space with high degree of freedom.

The purpose of this study is to present results of forecast of ranges for yen to US dollar exchange rate fluctuation in order to evaluate the performance of two algorithms: the original backpropagation (OBP), which is the most widely used algorithm, and the second algorithm (NBP), which is a proposed modification to the first one by the authors. The set of data consisted of economic and financial values that have already been calculated by the Bank of Japan and the Japanese Ministry of Planning and Finance. This data was available though the Nikkei Data Service and stretched from January, 1986, to the end of December, 1992. The results obtained show not only that NBP performs better than OBP since the former speeds up convergence time to a given error value, but also NBP shows a good generalization performance.

A method is presented for implementing a neural control system for controlling a piezopolymer bimorph flexible micro-actuator. Two neural controllers were constructed, both with an adaptive-type neural identifier and a learning-type direct or open-loop neural controller, focusing on the difference in learning speed between the adaptive and learning schemes. Simulated use of the proposed controllers to control a flexible micro-actuator showed that they can do so effectively. Experiments also demonstrated that a neural controller can be used to control a flexible micro-actuator.

In this paper an analog CMOS implementation of Approximate Identity Neural Networks is suggested. In particular a one-input one-output Neural Network with 6 neurons has been designed and fabricated with a 2µm CMOS technology. Due to the small area occupied the circuit proposed for the neuron is suited for the implementation of larger networks.

This paper obtains some new results about the existence, uniqueness, and global asymptotic stability of the equilibrium of a nonlinear continuous neural network, under a sufficient condition weaker than ones presented in the literature. The avobe obtained results can also imply the existing ones about avsolute stability of nonlinear continuous neural networks

In this paper, we apply the method of relating learning to hypothesis testing [6] to study average generalization performance of concept learning from noisy random training examples. A striking aspect of the method is that a learning problem with a so-called ill-disposed learning algorithm can equivalently be reduced to a simple one, and for this simple problem, even though a direct and exact calculation of the learning curves might still be impossible, a thorough empirical study can easily be performed. One of the main advantages of using the illdisposed algorithm is that it well models lower quality learning in real situations, and hence the result can provide useful implications as far as reliable generalization is concerned. We provide empirical formulas for the learning curves by simple functions of the noise rate and the sample size from a thorough empirical study, which smoothly incorporates the results from noise-free analysis and are quite accurate and adequate for practical applications when the noise rate is relatively small. The resulting learning curve bounds are directly related to the number of system weights and are not pessimistic in practice, and apply to learning settings not necessarily within the Bayesian framework.

We analyze additive effects of nonlinear dynamics for conbinatorial optimization. We apply chaotic time series as noise sequence to neural networks for 10-city and 20-city traveling salesman problems and compare the performance with stochastic processes, such as Gaussian random numbers, uniform random numbers, 1/fα noise and surrogate data sets which preserve several statistics of the original chaotic data. In result, it is shown that not only chaotic noise but also surrogates with similar autocorrelation as chaotic noise exhibit high solving abilities. It is also suggested that since temporal structure of chaotic noise characterized by autocorrelation affects abilities for combinatorial optimization problems, effects of chaotic sequence as additive noise for escaping from undesirable local minima in case of solving combinatorial optimization problems can be replaced by stochastic noise with similar autocorrelation.

Recurrent neural networks have the potential of performing parallel computation for associative memory and optimization, which is realized by the electronic implementation of neural networks in VLSI technology. Since the time delays in real electronic implementation of neural networks are unavoidably encountered and they can cause systems to oscillate, it is thus practically important to investigate the qualitative properties of neural networks with time delays. In this paper, a class of sufficient conditions is obtained, under which neural networks are globally asymptotically stable independent of time delays.

In this paper, a multiuser receiver using a Hopfield network (Hopfield network receiver) for asynchronous codedivision multiple-access systems is proposed. We derive a novel likelihood function for the optimum demodulation of a data subsequence whose length is far shorter than that of the entire transmitted data sequence. It is shown that a novel Hopfield network receiver can be derived by exploiting the likelihood function, and the derived receiver leads to a low complexity receiver. The structure of the proposed receiver consists of a bank of correlators and a Hopfield network where the number of units is proportional to both the number of users and the length of a data sequence demodulated at a time. Computer simulation results are presented to compare the performance of the proposed receiver with those of the conventional multiuser detectors. It is shown that the proposed receiver significantly outperforms the correlation receiver, decorrelating detector and multistage detector, and provides suboptimum performnace.

This paper describes an adaptive neural vector quantization algorithm with a deleting approach of weight (reference) vectors. We call the algorithm an adaptive learning and self-deleting algorithm. At the beginning, we introduce an improved topological neighborhood and an adaptive vector quantization algorithm with little depending on initial values of weight vectors. Then we present the adaptive learning and self-deleting algorithm. The algorithm is represented as the following descriptions: At first, many weight vectors are prepared, and the algorithm is processed with Kohonen's self-organizing feature map. Next, weight vectors are deleted sequentially to the fixed number of them, and the algorithm processed with competitive learning. At the end, we discuss algorithms with neighborhood relations compared with the proposed one. The proposed algorithm is also good in the case of a poor initialization of weight vectors. Experimental results are given to show the effectiveness of the proposed algorithm.

In the present communication we propose the application of unsupervised Artificial Neural Networks (ANN) to solve general ill-posed problems and particularly we apply them to the the estimation of the direction of arrival (DOA) of VLF/ELF radio waves. We use the wave distribution method which consists in the reconstruction of the energy distribution of magnetospheric VLF/ELF waves at the ionospheric base from observations of the wave's electromagnetic field on the ground. The present application is similar to a number of computerized tomography and image enhancement problems and the proposed algorithm can be straightforwardly extended to other applications in which observations are linearly related to unknowns. Then, we have proven the applicability and also we indicate the superiority of the ANN to the conventional methods to handle this kind of problems.

In the paper a new type of Multilayer Perceptron, developed in Quaternion Algebra, is adopted to realize short-time prediction of chaotic time series. The new introduced neural structure, based on MLP and developed in the hypercomplex quaternion algebra (HMLP) allows accurate results with a decreased network complexity with respect to the real MLP. The short term prediction of various chaotic circuits and systems has been performed, with particular emphasys to the Chua's circuit, the Saito's circuit with hyperchaotic behaviour and the Lorenz system. The accuracy of the prediction is evaluated through a correlation index between the actual predicted terms of the time series. A comparison of the performance obtained with both the real MLP and the hypercomplex one is also reported.

In this article, a new analogic CNN algorithm to extract features of postage stamps in gray-scale images Is introduced. The Gradient Controlled Diffusion method plays an important role in the approach. In our algorithm, it is used for smoothing and separating Arabic figures drawn with a color which is similar to the background color. We extract Arabic figures in postage stamps by combining Gradient Controlled Diffusion with nearest neighbor linear CNN template and logic operations. Applying the feature extraction algorithm to different test images it has been verified that it is also effective in complex segmentation problems

In this paper, after the introduction of the definition of State Controlled Cellular Neural Networks (SC-CNNs), it is shown that they are able to generate complex dynamics of circuits showing strange behaviour. Theoretical propoitions are presented to fix the templates of the SC-CNNs in such a way as to exactly match the dynamic behaviour of the circuits considered. The easy and cheap implementation of the proposed SC-CNN devices is illustrated and a gallery of experimentally obtained strange attractors are shown to confirm the practical suitability of the outlined strategy.

Electronic systems are progressively replacing mechanical devices or human operation for identifying people or objects in everyday-life applications. Especially, the contactless identification systems available today have several advantages, but they cannot handle easily several simultaneously present items. This paper describes a solution to this problem, based on blind source separation techniques. The effectiveness of this approach is experimentally demonstrated, especially by using a real-time DSP-based implementation of the proposed system.

Gray scale images are represented by recurrent neural subnetworks which together with a competition layer create an associative memory. The single recurrent subnetwork Ni implements a stochastic nonlinear fractal operator Fi, constructed for the given image fi. We show that under realstic assumptions F has a unique attractor which is located in the vicinity of the original image. Therefore one subnetwork represents one original image. The associative recall is implemented in two stages. Firstly, the competition layer finds the most invariant subnetwork for the given input noisy image g. Next, the selected recurrent subnetwork in few (5-10) global iterations produces high quality approximation of the original image. The degree of invariance for the subnetwork Ni on the inprt g is measured by a norm ||g-Fi(g)||. We have experimentally verified that associative recall for images of natural scenes with pixel values in [0, 255] is successful even when Gaussian noise has the standard deviation σ as large as 500. Moreover, the norm, computed only on 10% of pixels chosen randomly from images still successfuly recalls a close approximation of original image. Comparing to Amari-Hopfield associative memory, our solution has no spurious states, is less sensitive to noise, and its network complexity is significantly lower. However, for each new stored image a new subnetwork must be added.

This paper describes some dynamical properties of higher order neural networks with decreasing energy functions. First, we will show that for any symmetric higher order neural network which permits only one element to transit at each step, there are only periodic sequences with the length 1. Further, it will be shown that for any higher order neural network, with decreasing energy functions, which permits all elements to transit at each step, there does not exist any periodic sequence with the length being over k + 1, where k is the order of the network. Lastly, we will give a characterization for higher order neural networks, with the order 2 and a decreasing energy function each, which permit plural elements to transit at each step and have periodic sequences only with the lengh 1.

In this paper, we propose a modified Hamming network which contains less connection numbers and faster convergence speed. Besides, the real weight of subnet can also be transformed into integer weight. As so it is suitable for the hardware implementation of VLSI.

We develop a simulation environment for designing and examining a neural network model at the network level. The aim of our research is to enable researchers investigating neural network connective models to save time by being equipped with a graphical user interface and database of the network models. This environment consists of three parts: (1) the kernel of the simulation system, (2) NNDBMS (Neural Networks DataBase Management System), and (3) a system for displaying simulation results in various ways.