Search

The present authors recently reported synchronization phenomena in a star connection of several oscillators with identical characteristics through resistors. In such a system there occurs synchronization of the oscillators with deviated phases. Hence, extremely many phase conditions can be observed. The authors have also proposed a coupled oscillator network by extending this system. In the present network, many different phase patterns are observed. Based on its configuration, this network is a candidate for application as a cellular neural network (CNN). When this network is applied as a CNN, control parameters that characterize the behavior of the network are needed. In the present paper, the coupling resistors are recognized as such parameters. This paper examines how the variations of the coupling resistors affect the relationship between the initial state and the observed steady-state pattern. © 2003 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 87(1): 29–37, 2004; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.10102 [ABSTRACT FROM AUTHOR]

This paper proposes a back propagation method using a periodic activation function for a hierarchical neural network and examines its learning capability. As a specific learning problem, the paper addresses the n-bit parity problem and conducts a learning process by using a three-layer neural network. The computer simulation results indicate an increased learning coefficient range achieving a 100% learning success rate with an improved speed of learning compared with a neural network model using a monotonic activation function. In addition, the new method lessens the influence of the initial value that is set at the beginning of a learning process, and reduces the degree of dependency on parameters. © 2002 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 86(3): 11–19, 2003; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.1136 [ABSTRACT FROM AUTHOR]

This paper concerns our system of detecting ALS patients' intentional winking. Our previous paper was unfortified for patients who stirred somewhat. The present paper describes how to correct this problem. This is done by using four exclusive neural networks (NN). (1) The horizontal neural network detects the horizontal outlines of the upper and lower eyelids, if the eye is present in the input data of NN. (2) The vertical neural network detects that the two horizontal outlines are vertically close to each other. (3) The integrated neural network detects whether or not the eye has closed. (4) The time-delayed neural network detects intentional winking corresponding to the time that the eye has been closed. It is demonstrated that our system can detect intentional winking, even if a part of the eyelid in the input data wanes. © 2000 Scripta Technica, Electron Comm Jpn Pt 3, 83(9): 84–93, 2000 [ABSTRACT FROM AUTHOR]

This paper proposes a quantizer based on the Simple Hysteresis Network (SHN). The SHN is a hysteresis neural network in which all of the mutually coupled loads (neurons having hysteresis characteristics) are simplified. An A-D converter proposed by J. J. Hopfield uses neural networks, but this is highly initial-value-dependent so that the output for an arbitrary input is not necessarily satisfactory. The quantizer proposed in this paper guarantees the best output for an arbitrary analogue input. The proposed quantizer requires N wiring lines to obtain N bits of the output. This is significantly less than the A-D converter proposed by Hopfield, and practical circuits can easily be designed. © 1999 Scripta Technica, Electron Comm Jpn Pt 3, 83(4): 1–8, 2000 [ABSTRACT FROM AUTHOR]

This paper shows that two complex coefficient transfer functions can be realized by a single hypercomplex coefficient network, utilizing all four outputs of the hypercomplex network. The realization is considered for the case where the two arbitrary complex coefficient transfer functions do and do not have a common denominator. It is shown that the transfer functions can he realized by the hypercomplex coefficient network of the same order in the former case and by the hypercomplex coefficient network with the half-order in the latter case. The realization of the transfer functions with the common denominator is interesting, but the general method to derive the transfer functions is not known. From such a viewpoint, this paper proposes the discrete-type Kautz approximation with the positive frequency as the domain of approximation, including the optimal pole determination algorithm. By this approach, two complex coefficient transfer functions with an n-th-order common denominator are derived from the two 2n-th-order real or complex coefficient transfer functions. The method can be considered as one that simultaneously approximates the amplitude and the phase. The proposed approximation method is an approximation only for the positive frequency domain, and the property of the analytic signal that the negative frequency component is zero can be utilized. As a result, the method proposed in this paper has the advantages that the degree of freedom in the filter design is enhanced and the sampling frequency can be halved. In addition, the network structure can be simplified. [ABSTRACT FROM AUTHOR]

This paper introduces a feedforward network with a new learning algorithm into which both a self-organized mechanism and a supervised training mechanism are implemented. In particular, from some experiments, it is seen that this network can realize sophisticated signal processing and analysis data with high-speed training ability. The network is divided into two layers, that is, a self-organized layer and a supervised layer. However, in the self-organized layer, the training data of the network also are used for realizing high precision. In addition, the supervised layer also is trained in order to realize quite high accuracy for data recognition. In other words, the training data are used for both layers. Accordingly, the total precision of the network is improved greatly. In addition, the high-speed organization and the parallel processing which are embedded in the original self-organization also are realized in the new network. In this paper, some comparisons in which conventional neural networks, i.e., multilayer perceptron and self-organization models, are used also are shown. In particular, as a new application of speech ARM A modelling, a new spectrum envelope estimation is explained by using this network. [ABSTRACT FROM AUTHOR]

This paper proposes a new network that evaluates the locations of targets, i.e., target characters, and recognizes them in an image mixed with other objects, i.e., noncharacters. This network consists of a self-organized clustering layer, a simple perceptron layer, and a controller. The purpose of the self-organized clustering is to evaluate the stochastic characteristics of an input data set; the self-organization is executed by calculating the distance between stochastic characteristics in this layer and the stochastic characteristics of an input vector. The simple perceptron layer determines the location of a target character and simultaneously recognizes its character by linearly dividing all outputs given from the self-organized clustering layer. In the proposed system, the location identification and its recognition of a target character can be adaptively performed by using two small different comparative networks. The location identification and its recognition of real target characters, which are obtained in printed boards, are also shown in this paper. [ABSTRACT FROM AUTHOR]

The cellular neural network (CNN) is composed of a planar placement of cells which consists of a (piecewise-linear) nonlinear element and controlled current sources. It features a simple structure which is close to that of the retina and is expected to be utilized in pattern recognition and image processing. In CNN, each cell is connected to the neighborhood cells by the same pattern, and by adjusting the connection pattern, CNN with various functions can be designed. The connection pattern is called the cloning template. It is very important in the development of new CNN to establish the design method for the cloning template. In this paper, the operational characteristics are assumed so that the output for the normative input satisfies the specified conditions. The constraint is derived from the specification and the template is designed so that the corresponding cost function is minimized. The simplex method is used as the optimization technique, which features a simple algorithm. As application examples, noise-remover CNN as well as the maze-tracing CNN are designed and satisfactory results are obtained. The design method is reported in this paper. [ABSTRACT FROM AUTHOR]

The information storage algorithm of a Hopfield-type network is a kind of self-correlation learning. This algorithm needs the condition in which vectors of learning patterns cross at right angles to each other. This paper proposes n method to make this condition easier. To this purpose, threshold is used. An investigation of how to decide quantity is included. This paper discusses a network with units whose state is 1 or 0. Fixing method of threshold has not been proposed before to this type of network. It is, therefore, proposed in this research; the goal is to facilitate learning patterns. [ABSTRACT FROM AUTHOR]

This paper defines the degree of nonlinearity as a measure for the nonlinearity of the system. A method for estimating the nonlinearity is proposed based on the input and the output time-series signals under the actual condition that the additive observation noise is included. The degree of nonlinearity of the system takes a value between 0 and 1. It approaches 1 when the part of the variational power that carmot be represented by the linear combination of the input increases in the output variation depending on 0-input in the linear system. A multilayer perceptron is introduced as a parametric function which represents the wide class of nonlinear functions needed in the estimation. An example is shown of the family of memoryless and nonlinear systems where the degree of nonlinearity changes from 0 to 1 by the change of the system parameter. An example also is shown of the system with a finite memory having a degree of nonlinearity of 1. By a computer simulation, the validity of the nonlinearity estimation proposed in this paper is demonstrated. The method will be applied effectively to the modeling of the system based on the observed input and output signals. [ABSTRACT FROM AUTHOR]

This paper proposes a fast learning algorithm of neural networks and evaluates the performances of adaptive equalizers using neural networks trained by the proposed algorithm in a frequency-selective fading channel. The backpropagation (BP) algorithm which is used widely to train neural networks has a slow convergence rate because it is based on the gradient descent method. This paper presents a fast learning algorithm using the recursive least squares (RLS) algorithm which has a fast convergence rate as an adaptive algorithm for adaptive linear filters. In the proposed algorithm, the sum of the squared error between the actual total input and the desired total input is used as the cost function to apply the RLS algorithm. A simulation result on the exclusive-OR problem indicates that the proposed algorithm is about 8.8 times faster than the BP for the number of iterations required to converge. Recently, there has been interest in adaptive equalizers as an application field of neural networks. However, the performance of an adaptive equalizer using a neural network in a frequency-selective fading channel which is observed in land mobile communications has never been evaluated. Therefore, in this paper, the performances of adaptive equalizers using neural networks trained by the proposed algorithm in a frequency-selective fading channel are evaluated. Especially, an adaptive equalizer using the selectively unsupervised learning neural network proposed by the authors is considered. The adaptive equalizer can reject the false learning by carrying out learning selectively. It is shown that the adaptive equalizer is superior to the conventional one and the one using the conventional neural network. [ABSTRACT FROM AUTHOR]

This paper considers on-line learning of neural networks for an inverse problem. A neural network must learn an inverse system of a static unknown system on the basis of an evaluation function that is a squared error between an output of the unknown system and its desired value. Usual learning methods need information about the sign of the sensitivity of the unknown system. This paper proposes learning rules that do not require the information about that sign. These rules utilize a perturbation signal. Rules with the perturbation on an input of the unknown system and a rule with it on each weight of the neural network are presented. The usefulness of these learning rules is confirmed via some numerical simulations. [ABSTRACT FROM AUTHOR]

This paper proposes recognition methods that use hierarchical spectral dynamic features extracted over multiple time lengths. The effectiveness of these features in phoneme recognition and isolated word recognition is shown. The dynamic features are extracted for time sequences of both cepstral coefficients and logarithmic energy, and are combined with instantaneous cepstral coefficients. Input speech Is quantized by word- or phoneme-specific codebooks created by clustering a set of vectors consisting of these features for each word or phoneme. Speaker-independent isolated word recognition experiments are performed using a vocabulary of 100 Japanese words. When VQ distortion is used for word identification, the high recognition accuracy of 96 percent is achieved, and when VQ distortion is used for preprocessing, the number of word candidates for each input utterance is reduced to 1 percent of the vocabulary without increasing the error rate. Phoneme recognition experiments are performed far the /b/, /d/ and /g/ consonants in a large vocabulary of isolated words uttered by one male speaker. Using the proposed recognition method, the high recognition accuracy of 99 percent, which is similar to the accuracies obtained by conventional HMM or neural network approaches, is obtained. This paper also compares multiple codebook and single code-book methods. [ABSTRACT FROM AUTHOR]

OFDM communication system has recently been attracting attention because of its features of high spectrum efficiency and robustness in a multipath environment. The OFDM signal often exhibits a high peak power. It is important to reduce the peak power, since communication quality is degraded and unnecessary outband power radiation is produced when such a signal is amplified by a nonlinear amplifier. This paper considers the peak power suppression problem in an oversampled OFDM signal and presents a solution using a mutually coupled neural network. The peak power suppression performance of the proposed method and the accompanying reduction of the outband radiation power and the bit error rate are evaluated by numerical experiments. Neuro-hardware which can handle the peak power suppression process in real time is experimentally constructed, and its performance is evaluated. © 2007 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 90(8): 54– 62, 2007; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.20285 [ABSTRACT FROM AUTHOR]

This paper proposes a new error backpropagation method (DBP) for a three-layered neural network containing a nondifferentiable binary output unit. In contrast to the conventional simple perceptron, in which the teacher signal is given only to the output layer, in the DBP method the teacher signal is also given to the middle layer so that the output error is decreased. Consequently, it is possible in the DBP method to correct the coupling weights in both the lower layer and the upper layer. This makes it easy to construct a network composed only of binary output units, which results in high-speed operation and is suitable for hardware implementation. When the DBP method is applied to linearly inseparable tasks such as XORing, the learning performance is greatly improved compared to learning by the simple perceptron, and almost the same learning performance as the conventional BP is obtained. © 2007 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 90(5): 40– 49, 2007; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.20318 [ABSTRACT FROM AUTHOR]

For the purpose of providing disaster prevention information, such as information concerning mudslides, flooding, and rock avalanches, it is considered important to predict local sudden precipitation accurately. However, no solution to this problem has been achieved. This paper considers the problem and proposes a precipitation estimation system using a support vector machine and a neural network. It is noted that precipitation is a very complex nonlinear event, and a multistage estimation system is proposed to deal with its complexity. The proposed system has the feature that the Euclidean distance between the support vector in the support vector machine and the evaluation data in the original feature space is used to estimate the precipitation. Computer experiments indicate that the proposed method gives a mean-square error of the evaluated data from the actually measured value which is better than the best result reported to date. Thus, the usefulness of the proposed estimation system is demonstrated. © 2005 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 89(3): 38–47, 2006; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.20226 [ABSTRACT FROM AUTHOR]

It is known that rats with hippocampal lesions exhibit abnormal spatial navigation behavior. In order to analyze the role of the hippocampus of the rat in performing spatial navigation tasks, this paper presents a hippocampus–basal ganglia model. The hippocampus in this model integrates multiple visual information inputs from the visual association area, and the basal ganglia perform reinforcement learning in order to acquire a stochastic mapping from the information obtained from the visual association area and the hippocampus to the output action. Assuming milk maze tasks with visible and invisible platforms, numerical simulations are performed using the proposed model. The result is that both normal rats and rats with hippocampal lesions can perform the milk maze task with a visible platform. In the milk maze task with an invisible platform, on the other hand, rats with hippocampal lesions require a longer time than normal rats to arrive at the platform. These results agree with the experimental results. © 2004 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 87(10): 66–80, 2004; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.20109 [ABSTRACT FROM AUTHOR]

In this paper, a new adaptive equalizer that adjusts the nonlinear filter without training signals is proposed and its performance is evaluated. In the proposed equalizer, the nonlinear filter has constraints, while learning is carried out by means of the output energy minimization algorithm. Since the proposed equalizer has nonlinearity, it has a better bit error rate than the linear equalizer. Further, since no training signal is needed, it is to be expected that the transmission efficiency is superior to that in an equalizer with training signals. In order to evaluate the performance limitations of the proposed equalizer, an optimum equalizer with constrained filter input is derived. By means of simulations, a performance evaluation is carried out and the effectiveness of the proposed equalizer is demonstrated. Further, in order to improve the performance of the proposed equalizer for a linear communication channel, an intersymbol interference canceler and an algorithm for correcting the error are proposed and their effectiveness is demonstrated. © 2003 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 87(3): 17–29, 2004; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.10145 [ABSTRACT FROM AUTHOR]

When the characteristics of non-Gaussian time series, such as biological signals, are described, higher-order spectra such as the bispectrum must be referenced in addition to the power spectrum. To obtain estimates having few statistical fluctuations, parametric estimation methods of these spectra should be established. In recent years, a simple discretization method (SDM) has been proposed as a numerical solution based on iterated integral transform of the stationary joint density probability function and the higher-order spectra of nonlinear autoregressive models like a neural network autoregressive model. However, the SDM requires an amount of calculation memory that increases exponentially with respect to the lag order of corresponding models. In order to overcome this drawback, the tensor product expansion method (TPEM) where the joint probability density function is expanded as the sum of single-variable functions was proposed. This method can suppress the increase in the required memory capacity to a linear increase. A numerical example has shown where the TPEM calculated the bispectrum more accurately than the SDM with an equivalent amount of memory. However, the relationship between the performance of tensor product expansion and the dependence properties of the target time series has not as yet been investigated. In this paper, we study the property dependence of the time series in the TPEM and point out the problem of worsening calculation accuracy for a time series in which values at each time instant highly depend on each other. As the solution, we propose a new method that introduces a linear coordinate transform in state space where the joint probability density function undergoes the tensor product expansion. © 2003 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 87(1): 19–28, 2004; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.10103 [ABSTRACT FROM AUTHOR]

The 1/f spectra that appear in electroencephalogram data are among the characteristics that appear in the wide range of self-organizing phenomena encompassing both physical and sociological phenomena. The fact that they appear across the various objects suggests that the underlying cause of the 1/f spectra can be elucidated by using a simplified model that does not depend on the detailed structure of the objects. This paper, treating the electroencephalogram as a macro-quantity of neuron activities, demonstrates the existence of a self-similar solution through the spatial-temporal coarse-graining of a neural network, which is a simplified model of the brain. The self-similar solution in fact leads to the derivation of a 1/f spectrum. © 2003 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 87(1): 72–81, 2004; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.10016 [ABSTRACT FROM AUTHOR]

This paper describes the construction of a system for generating linguistic expressions explaining circumstances related to people or objects that appear in multiple pictures as a means of modeling one form of human intellectual information processing. This system outputs linguistic expressions explaining the emotions of a human object, positional relationship between objects, apparent behavior of an animate object, or relationships connecting pictures based on objective information such as the facial expression, position, or size of an object. To implement the system, soft computing techniques such as neural networks, fuzzy reasoning, and case-based reasoning are used. Neural networks are used to recognize the degree of emotion from the facial expression of a human object, and fuzzy reasoning is used to infer the degree to which an animate object can discern another object. Case-based reasoning is used to explain the apparent behavior of an object and relationships connecting pictures. Fuzzy sets are used to convert the information that is obtained to the linguistic expressions that are output. Finally, the effectiveness of the system was confirmed by using the constructed system to perform simulations and evaluation experiments. © 2003 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 86(12): 38–53, 2003; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.10141 [ABSTRACT FROM AUTHOR]

High-dimensional neural networks, such as complex neural networks and three-dimensional neural networks, have been proposed. They are capable of learning high-dimensional motions, which cannot be done by real-valued neural networks. In this paper, high-dimensional neurons are defined as vectors, and certain high-dimensional neural networks, such as complex neural networks, quaternary neural networks, and three-dimensional exterior neural networks, are unified in terms of a vector representation. It is shown that all of these types have the common feature of alternating matrices. The inverse quaternion neural network has the ability to learn rotations. In addition, it is shown that the proposed model cannot represent all rotations in neural networks that are higher than three-dimensional, and that the inverse quaternary neural network has a partial ability to learn rotations. © 2003 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 86(11): 38–45, 2003; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.10072 [ABSTRACT FROM AUTHOR]

Many studies of the neural network with chaotic output have been presented. In most of these, however, a neuron model or a network model is specially prepared to realize the chaotic response, and few studies have been based on the conventional model. This paper considers a neural network based on the conventional neuron model, such as is used in the Hopfield and other neural networks, and discusses the oscillation phenomenon. It is shown theoretically and experimentally that the behavior of the network composed of two neurons is either convergence or periodic oscillation. Then the network composed of three neurons is examined experimentally, and it is shown that convergence, periodic oscillation, quasi-periodic oscillation, or chaos can occur. In other words, it is shown that the minimum configuration of a neural network based on the conventional neuron model that can generate chaos is a network composed of three neurons. © 2003 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 86(7): 62–71, 2003; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.1152 [ABSTRACT FROM AUTHOR]

When an episodic memory is to be transferred from short-term memory to long-term memory, there must be a buffer mechanism that temporarily stores the content of the short-term memory as a mid-term memory. It is considered that the hippocampus plays that role. In composing an episodic memory, there must be not only the association of different events without a time factor (spatial pattern), but also the association of different events containing time factors (time sequence pattern). This paper proposes a hippocampus neural network model, based on anatomical and physiological knowledge. The model features the following points. (i) By applying a learning rule where the input history is considered in the hippocampus CA3, which is regarded to have the associative recall function, it is made possible for the model to learn the time sequence pattern. (ii) The orthogonal coding function is provided in the gyrus dentatus, and the decoding function is provided in the hippocampus CA1. It is expected that the memory capacity of the hippocampus will be enlarged by providing the orthogonal coding function. It is shown by a numerical simulation that the mid-term memory can learn spatial and time sequence patterns, and that associative recall of the memory contents can be realized. This result implies that the proposed hippocampus model can retain episodic memory as a mid-term memory. The obtained results agree with physiological data, and demonstrate the effectiveness of the proposed model. © 2003 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 86(6): 86–97, 2003; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.1010 [ABSTRACT FROM AUTHOR]

When direction-of-arrival estimation methods based on eigenvalue expansion such as MUSIC and ESPRIT are applied to an ideal sensor array, high resolution can be attained. However, if there exist array sensor errors due to sensor position errors and degradation of the phase shifters, significant errors may arise in the estimation results. Also, in order to obtain sufficient estimation accuracy, a relatively high signal-to-noise ratio is required in the array input. In this paper, the Hopfield neural network is employed for direction-of-arrival estimation and a direction-of-arrival estimation method that is robust to array errors by virtue of using training signals is proposed. By means of computer simulation, the method is compared with that of MUSIC and its effectiveness is verified. © 2003 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 86(6): 19–28, 2003; Pub-lished online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.10062 [ABSTRACT FROM AUTHOR]

In the correlation-type associative memory model it is known that the storage capacity per synapse increases when synapses are pruned. In that case, however, the storage capacity of the whole network is decreased. In order to solve this difficulty, the authors introduce the synapse with delay, and propose a scheme in which the coupling rate of the synapse is reduced while retaining the number of synapses. As an example of such an approach, this paper discusses a discrete-time synchronous model with delayed coupling and synaptic pruning. The dynamics of the model is investigated theoretically by statistical neurodynamics, and the storage capacity is analyzed quantitatively. An analysis is performed for two types of synaptic pruning, random and systematic. In either case, it is seen that the storage capacity increases when the number of delay stages is increased while reducing the synaptic coupling rate, provided that the total number of synapses is kept constant. The increasing tendency is more marked in systematic pruning than in random pruning. These results were verified by comparing the results to those of computer simulation. The significance of the theoretical findings and the excessive generation and pruning of the synapses in the brain are also discussed from the viewpoint of the theory of computation. © 2003 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 86(6): 48–58, 2003; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.10037 [ABSTRACT FROM AUTHOR]

Detection of the pitch frequency as the basic characteristic of the voice is one of the most important research subjects in voice analysis and synthesis. In this paper we study, in comparison with the cepstrum method and the modified correlation method, how the BPFP method using bandpass filter pairs and the BPFP-NN method combining the BPFP banks and the neural network (NN) proposed by the authors perform in the presence of pitch disturbance, waveform variation, and added noise when applied to voiced/unvoiced (U/V) detection and pitch extraction. According to the experimental results, the BPFP method and the BPFP-NN method exhibit more stable performance and more effective pitch extraction with U/V detection than the typical cepstrum method and modified correlation method, and the effect is particularly evident at low frequencies due to the logarithmic expression of the center frequency spacing of the BPFP banks. © 2003 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 86(5): 24–35, 2003; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.10045 [ABSTRACT FROM AUTHOR]

Instead of the conventional static memory model based on an associative memory neural network, this paper proposes a dynamic memory model that is extended to the input-output system in order to allow interaction with external information. Based on the time correlation between the input signal and the network output, we analyze and investigate the network response to stored, nonstored, and mixed stored pattern signals, and the parameter dependencies inherent in the model. The possibility of a new mechanism, the chaotic dynamic model, which discriminates and identifies pattern signal information by combining the dynamic memory model with its extension is shown to be possible. © 2003 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 86(5): 54–64, 2003; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.10022 [ABSTRACT FROM AUTHOR]

In this paper we analyze observation data which consist of environmental factors around a creature as the explanatory variables and a population number as the explained variable. We choose data on the firefly. To analyze such data we propose a system incorporating probabilistic neural networks which can acquire an unknown nonlinear mapping from the explanatory variables to the explained variable by using learning. The proposed system can estimate the effect of the explanatory variables on the explained variable, that is, it can solve the inverse problem. To realize the desired environment for the selected creature, we show that our proposed system can suggest an adequate strategy for the controllable explanatory variables. © 2002 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 85(11): 18–25, 2002; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.10002 [ABSTRACT FROM AUTHOR]

The radio network is a large-scale communication network in which a large number of stations are distributed over a wide area. Each station can communicate directly by radio packets only with neighborhood stations. Packet communication between stations which cannot communicate directly is realized by passing packets through multiple stations between them. For this purpose, when a packet communication request arises, the route for the packet and the transfer timing must be chosen so that the total transfer time is minimized and the efficiency of the radio network is maximized. It should be noted that the reliability, representing the probability that the required performance is achieved, fluctuates among the stations in the network due to the differences in the time of installation, the designed performance, and communication congestion. Consequently, this paper proposes a two-stage algorithm in which the route is assigned so that its reliability is larger than the required value. The proposed algorithm is composed of the iterative reliability satisfaction extraction procedure, in which a set of path candidates satisfying the reliability condition for the given transfer request, and the greedy neural network procedure, in which the route minimizing the cost is selected from the set of path candidates. The effectiveness of the proposed algorithm is demonstrated by simulation for networks with up to 500 stations. © 2002 Scripta Technica, Electron Comm Jpn Pt 3, 85(6): 63–73, 2002 [ABSTRACT FROM AUTHOR]

We employ statistical dynamics to study the convergence of the Wake–Sleep (W-S) algorithm, which is a learning algorithm for neural network models having hidden units. Although there have been several reports on the effectiveness of the W-S algorithm based on experimental methods, the theoretical side is not clear even for a simple network. In this paper, we investigate the dynamic characteristics of the W-S algorithm expressed by a single factor analysis problem, which is the simplest state setting. The advantage of our approach is the ability to quantitatively evaluate the effect that the learning coefficients have on the convergence, which is difficult when using other methods. The result was that the settings of the learning coefficients, particularly in the Sleep step, had a substantial effect on the convergence of the algorithm. © 2001 Scripta Technica, Electron Comm Jpn Pt 3, 85(1): 41–49, 2002 [ABSTRACT FROM AUTHOR]

In the enlargement of a digital image, it is required to estimate the high-frequency components, which were lost in sampling. As a method of meeting this requirement, the authors have proposed an image enlargement method using a high-resolution neural network (NN). The high-resolution neural network provided excellent results of image enlargement. There remained, however, the following two problems. (1) When the image is enlarged by the high-resolution NN, artifacts are produced in the smooth region of the enlarged image when the magnification ratio is increased as 4, 8, . . . . (2) There is a variation of approximately 10% in the accuracy of the enlarged result, depending on the kind of image used in the training of the high-resolution NN. In order to cope with such problems, this paper newly proposes the high-resolution multi-neural network (MNN), which is based on the local variance. The training procedure is shown. In the high-resolution MNN, two NN are used, corresponding to the region with a rapid change (high local variance) and the region requiring a smooth interpolation (low local variance), in the image for which resolution improvement is required. The weighted sum of the two NN outputs is defined as the result of enlargement. It is shown through application examples that the two problems in the high-resolution NN are remedied by using the high-resolution MNN. © 2001 Scripta Technica, Electron Comm Jpn Pt 3, 84(11): 61–69, 2001 [ABSTRACT FROM AUTHOR]

If the inverse function x=f1(y) of the system y=f(x) with an input of x and an output of y is multivalued, learning of the inverse system by the neural network is difficult. Hence, in this paper, the mapping of (y,x)→x is learned by the neural network as the inverse system by making use of the extended generalized learning method. Next, the learning of the inverse system (many-valued function) is carried out by solving the obtained inverse system (nonlinear equation) by means of the successive substitution method after the desired output value yd is given so that the desired input value xd is derived. In this case, the learning problem of the neural network that satisfied the convergence condition of the successive substitution method is formulated as a satisfaction optimization problem. This is solved by the nondifferentiable optimization method. As a result of simulation experiment, it is confirmed that an inverse system of a one-to-multivalue correspondence is obtained. © 2001 Scripta Technica, Electron Comm Jpn Pt 3, 84(8): 57–66, 2001 [ABSTRACT FROM AUTHOR]

In this paper, we propose a hill-climbing learning method for Hopfield networks in which the energy of the network is intentionally raised in the weight space so that the network can escape from local minima. This learning method involves repeated updating of the Hopfield network in state space and modification of the weights in weight space if it settles into a local minimum, so that the energy is raised. Shifting of the global minimum during learning can be avoided by updating the state on the unlearned Hopfield network in state space, using equilibrium states obtained from learning as the initial state. Simulations using this learning method show that the Hopfield network can escape from a local minimum which depends on the initial state. A simulation experiment is first conducted with a two-variable Hopfield network, and the convergence of learning and the change of the domain from a local minimum to a global minimum are demonstrated. Then, using an example of the traveling-salesman problem, a simulation is performed using concrete numerical values, and the validity of the method is demonstrated. © 2001 Scripta Technica, Electron Comm Jpn Pt 3, 84(7): 28–40, 2001 [ABSTRACT FROM AUTHOR]

An important problem in image processing is noise removal that preserves edges in images degraded by white Gaussian noise. In this paper, we propose a data-dependent weighted mean filter using counter propagation networks. This method achieves high-level performance in the removal of white Gaussian noise while preserving the edges. Coefficients of the filter and features to change them are adjusted and decided by using a layered neural network and self-organizing network, respectively. This allows a robust filter to be obtained. Analysis of this network enables the derivation of a simpler filter structure. Using this fact, we show an example of analysis and simplification, and obtained a filter capable of processing at high speed by fewer computations. © 2000 Scripta Technica, Electron Comm Jpn Pt 3, 84(3): 42–51, 2001 [ABSTRACT FROM AUTHOR]

In this paper, a pulse-mode multilayer neural network with on-chip learning is proposed. The neuron unit uses voting circuit as a nonlinear adder to improve the nonlinear activation function. Moreover, the voting circuit is modified to have adjustable nonlinear characteristic. As the signal level is expressed by the frequency, synapse multipliers are realized by simple frequency converters. The back propagation algorithm is used for the on-chip learning. The proposed multilayer neural network is implemented on field programmable gate array (FPGA), and various experiments are conducted. The results show that the proposed neuron has adjustable nonlinear function. The learning capability of the proposed network is also verified by the experiments. © 2000 Scripta Technica, Electron Comm Jpn Pt 3, 84(1): 32–42, 2001 [ABSTRACT FROM AUTHOR]

In this paper, we propose a neural network algorithm that uses the expanded maximum neuron model to solve the channel assignment problem of cellular radio networks, which is an NP-complete combinatorial optimization problem. The channel assignment problem demands minimizing the total interference between the assigned channels needed to satisfy all of the communication needs. The proposed expanded maximum neuron model selects multiple neurons in descending order from the neuron inputs in each neuron group. As a result, the constraints will always be satisfied for the channel assignment problem. To improve the accuracy of the solution, neuron fixing, which is a heuristic technique used in the binary neuron model, a hill-climbing term, a shaking term, and an Omega function are introduced. The effectiveness of these additions to the expanded maximum neuron model algorithm is demonstrated. Simulations of benchmark problems demonstrate the superior performance of the proposed algorithm over conventional algorithms in finding the solution. © 2000 Scripta Technica, Electron Comm Jpn Pt 3, 83(11): 11–19, 2000 [ABSTRACT FROM AUTHOR]

Investigations into quantum computations have begun from the pioneering theoretical studies of Feynman, Deutsch, and others, and detailed studies have been done since the discovery of a quantum algorithm which can solve the problem of factorizing a large integer in polynomial time by Shor in 1994. Recently, the existence of nonalgorithmic quantum computations in microtubules inside a neural circuit has been debated, resulting in the proposal of the concept of the quantum neural computation theory, although detailed studies have not as yet been made. In this paper, in order to construct a new framework for describing the cohesiveness of the distribution and synthesis inherent in a neural network, a neural state is described quantum dynamically and a qubitlike neural network corresponding to the quantum circuit of quantum computations is studied. Specifically, a qubitlike neural network is constructed for a 3-bit quantum circuit, which is the minimum quantum logical gate describing all basic logical operations, and in this model we investigate how to determine circuit parameters by learning. © 2000 Scripta Technica, Electron Comm Jpn Pt 3, 83(10): 67–73, 2000 [ABSTRACT FROM AUTHOR]

When certain symmetry is assumed for an oscillator, an oscillation phenomenon caused by such symmetry can be observed. In this paper, a neural oscillator that generates an N-phase solution by a circulating structure in the system is studied. An E–I pair consisting of a mutual coupling of exciting and inhibitory neurons is considered. A system is proposed that consists of N twisted ring structure composed of the E–I pairs. The oscillation condition of this system is derived. It is found that a stable N-phase solution exists as one of the oscillation modes. Further, the various oscillation modes seen in the system with N = 3 are analyzed by means of the bifurcation diagram and a phase plane. © 2000 Scripta Technica, Electron Comm Jpn Pt 3, 83(9): 34–45, 2000 [ABSTRACT FROM AUTHOR]

Generally, to find the optimal release time for a software product, the parametric estimation values of the mean value function, which characterizes the software reliability growth model, are determined from the fault detection time data observed during the testing phase, and an analytical method is used to determine the time which minimizes the total software cost. On the other hand, in this paper we concentrate on the geometrical solution method for the optimal release problem based on the software reliability growth model, and the optimal release time is directly estimated from fault detection time data. Neural networks can be applied to the prediction of fault detection time data. It is shown that the optimal release time can be estimated with better accuracy by the proposed method using neural networks than by the conventional methods. © 2000 Scripta Technica, Electron Comm Jpn Pt 3, 83(8): 10–18, 2000 [ABSTRACT FROM AUTHOR]

In this paper, with a view to automatic selection of bottles in empty bottle recycling, an object recognition system based on the acoustic wave as an acquisition medium of the object information is described. The unique features of this system are: (1) the method can be used for optically transparent objects such as glass bottles, and (2) it can flexibly cope with shift and rotation of the objects. First, the acoustic Fourier transform (AFT) carrying out the Fourier transform of the object information by acoustic waves is presented. In this system, the power spectrum of the object information can be obtained instantaneously at the focal plane of the acoustic wave and the positional invariance of the object is automatically realized. Further, the rotational invariance of the object can also be realized easily by applying the polar coordinate transform to the power spectrum of the object and by considering the rotation as a shift in the rotating direction. For the recognition of the object, a three-layered neural network is used in which the input is the power spectrum with the positional and rotational invariance processes described above. By computer simulation, the design parameters of the experimental system are determined so that a simple experimental system is fabricated. The effectiveness of the present method is verified by the recognition experiments of 10 types of bottle objects specifically carried out for this verification. © 2000 Scripta Technica, Electron Comm Jpn Pt 3, 83(7): 42–52, 2000 [ABSTRACT FROM AUTHOR]

When a statistical estimation is carried out with a model that has layered parameters such as a neural network, the behavior of the generalization error and the optimum model design method are unknown, since unlike the regular model, the asymptotic behavior of the estimated parameters is not clear. In this paper, it is theoretically proven that the generalization error can be decreased if the Bayesian method is applied to a layered model. It is shown that the Bayesian method is effective as a learning method for a large-scale layered model. © 2000 Scripta Technica, Electron Comm Jpn Pt 3, 83(6): 95–106, 2000 [ABSTRACT FROM AUTHOR]

In a multicast packet switching system, multiple output ports can send out packets simultaneously by means of the packet copying function. In realizing such a switching system, the transmission pattern must be determined for successively arising transfer requests, so that the number of transmitted packets is maximized, while satisfying the constraints. This paper considers the one-shot scheduling problem formulated by Chen and colleagues, where all copies of the same packet are sent at the same time, and proposes a new solution method using a neural network. In the proposed solution method using the neural network, the binary neuron, which is suited to the digital techniques and hardware implementation, is employed together with the synchronous update scheme. In order to improve the convergence performance and the accuracy of the solution, an operational equation with an additional heuristic term and a coefficient setting procedure taking account of transfer requests are proposed. By simulations for cases of up to 1000 ports, it is shown that the proposed solution method using the neural network is better in deriving the solution than the conventional method. It is also shown that the considered problem is NP-hard. © 1999 Scripta Technica, Electron Comp Jpn Pt 3, 83(3): 12–21, 2000 [ABSTRACT FROM AUTHOR]

The authors previously reported on a variable logic circuit using a neural network (NN), in which logic operation was controlled by an analog voltage. In that design, the synaptic weights were determined by the error back-propagation (BP) algorithm. Then, logic outputs outside the specification were produced between logic outputs with Hamming distances of two or more. This paper proposes a design of NN in which no logic outside the specification is output, together with a method of calculating the synaptic weights. Based on the correspondence between the change of the output logic and the control voltage, the necessary number of hidden-layer neurons can be defined. The computation time for the synaptic weights is reduced. By constructing an NN circuit by the proposed method, it is verified that the desired operation is realized. © 1999 Scripta Technica, Electron Comm Jpn Pt 3, 83(1): 14–20, 2000 [ABSTRACT FROM AUTHOR]

This paper proposes a hypothetical theory called the “emergence hypothesis” that tries to unify the evolution of life and intelligence based on the superexponential law and computational universality. The superexponential law is an extension of the square-root law proposed by Kauffman, and its validity is examined by investigating human pattern recognition characteristics, neural networks, and molecular biological data. Computational universality, called “elemental universality,” is proved to be a necessary condition for the crystallization of order in the Kauffman networks, which means that the edge of chaos has potential computational universality. Elemental universality, decomposed into nonlinearity and negative feedback, can be regarded as a kind of natural law, which, however, may not belong to conventional materialistic physics. © 1999 Scripta Technica, Electron Comm Jpn Pt 3, 82(11): 1–8, 1999 [ABSTRACT FROM AUTHOR]

The graph partitioning problem is an NP hard problem of deriving the partitioning of each vertex such that the total sum of the edge weights among the groups is minimized and the total sum of the vertex weights in each group is less than the upper limit. In this paper, a neural network solution is proposed in which the binary neurons are used for the graph 2-partitioning problem. In the present neural network, an energy function that is applicable to graphs both with and without edge and vertex weights is defined. For improvement of solution accuracy, shaking terms are introduced into the operating equation. To evaluate the solution search capability of the present method, simulations are carried out for random graphs, together with the KL method proposed by Kernighan and Lin, and the FM method proposed by Fiduccia and Mattheyses. From the simulation results, it is shown that the solutions obtained by the present method are the best. © 1999 Scripta Technica, Electron Comm Jpn Pt 3, 82(12): 34–42, 1999 [ABSTRACT FROM AUTHOR]

In order to capture the characteristics of non-Gaussian time series observed in biological signals, it is necessary to study higher-order spectra, such as the bispectrum, in addition to the power spectrum. In order to obtain estimates with little statistical variations in the estimation of higher-order spectra, establishment of a parametric estimation method is desired. Recently, a method has been proposed to numerically derive the higher-order spectra of nonlinear autoregressive models, such as neural net autoregressive models, by means of the repetitive integral transform. However, in efforts to describe the probability density function simply in a discrete manner, there is the problem of an exponentially increasing number of data points to be computed with increasing lag order. In this paper, in order to resolve this problem, a numerical method is proposed for description of the probability density function and integral kernel by means of a tensor product expansion. The tensor product expansion is a representation in terms of the sum of the tensor products of characteristic one-variable functions depending on the given multivariable function. There is a possibility of reducing the memory capacity needed for the calculations. Numerical examples of accurate calculations with fewer data points in the proposed method than in the conventional method are presented. Finally, an example of parametric estimation of the power spectrum and bispectrum of Wolf's sunspot numbers by the proposed method is given. © 1999 Scripta Technica, Electron Comm Jpn Pt 3, 82(11): 47–57, 1999 [ABSTRACT FROM AUTHOR]

The recognition of human facial expressions has been expected to have applications in various fields, such as psychology, engineering, and so forth, and many techniques have been proposed to date. However, most of the techniques are based on the motions of local feature-bearing blocks, such as the eyes and mouth, which are supposed to be closely associated with facial expressions. It is thus required to segment these blocks from the face image and to track their motions in real-time applications, resulting in high complexity. Furthermore, not all information on the motions of the facial expressions has been utilized. In this paper, a new recognition technique is proposed, which uses the 2D DCT of the entire facial image and a neural network. The segmentation of local feature-bearing blocks is not needed. The differences of the DCT coefficients in the lower frequency area between neutral and expression-bearing images are given to a neural network to realize a mapping into the facial expression space. The new technique has been applied to a database of normalized facial images of 60 persons, with images of 40 persons used for network training and the remaining images for testing. The recognition rate is 100% for the training images. A maximum recognition rate of 95% is achieved for the testing images. © 1999 Scripta Technica, Electron Comm Jpn Pt 3, 82(7): 1–11, 1999 [ABSTRACT FROM AUTHOR]

This paper discusses the realization of an adaptive weighted median filter which can eliminate impulsive noise while preserving the original signal. We propose a method in which the weights are determined by a weight controller composed of a counterpropagation network, and the learning algorithm is investigated. The weight controller executes a kind of pattern matching, in which the inputs are classified according to their characteristics and the corresponding weight masks are assigned. By this technique, the preservation of the original signal is improved. Since a neural network is employed, the class of the input signal and the corresponding mask can be determined by learning, which lead to a satisfactory result. The effectiveness of the proposed method is demonstrated with an actual image processing example. © 1999 Scripta Technica, Electron Comm Jpn Pt 3, 82(6): 46–54, 1999 [ABSTRACT FROM AUTHOR]

The enlargement of a digital image is a process of narrowing the sampling interval. Since the image before the enlargement and that after the enlargement have different Nyquist frequencies, it is necessary to predict or estimate the high-frequency components that are lost in the image before processing and to complement these components in the enlargement process. It is noted that the image can be represented as the sum of Gaussian components and Laplacian components in a hierarchical structure (pyramid representation). There exists a strong correlation between the Laplacian images of the layers. Based on that property, Greenspan and Anderson have proposed a nonlinear procedure where the unknown Laplacian image of high resolution (i.e., containing the lost high-frequency components) is estimated from the Laplacian image of low resolution. This paper notes that the Greenspan procedure can be network-structured and proposes an enlargement method for digital images using a neural network (NN) for improving the resolution, based on generalizing the network. It is shown that the enlarged NN can be constructed almost independently of the kind of training image or of the hierarchy (resolution). It is also shown that a better enlarged image can be derived compared with Greenspan's method. In other words, the effectiveness and practical usefulness of the proposed method are demonstrated. © 1998 Scripta Technica, Electron Comm Jpn Pt 3, 82(1): 19–29, 1999 [ABSTRACT FROM AUTHOR]