TANAKA MAMORU

The covariance structure analysis has the advantage that make it visible to phenomena that cannot be expressed numerically. The result that using software of the covariance structure analysis is depending on an initial value and the model dependencyHere, it proposes a new algorithm that introduces the back Euler method. By using a new algorithm, it is an aim to weaken an initial value and the model dependency. In this research, the domination of a new algorithm was shown by analyzing the result of the questionnaire in the covariance structure analysis. The comparison that uses SDP method and free software R is being researched.

Covariance structure analysis is a multivariate analysis technique for appearing in the latter half of the 1960's. It to be a statistical approach to understand the social phenomenon and natural phenomena from introducing a potential variable that cannot do the direct detection, and identifying the causal relations between the potential variable and the observation variable, and enhancing of the factor analysis and the multiple regression analysis (path analysis). It has been applied to fields of a sick investigating the cause and the marketings etc. such as statistics, psychology, and depression now. It is possible to use in a very many topics field like this, and statistics and the analysis methods with a lot of advantages. In this thesis, a new study method by a new algorithm different from the current covariance structure analysis was applied to the frequency response data of the transfer function, it simulated, and it evaluated it.

LIFN(Leaky Integrated-and-Fire Neuron)is one of the neural circuit models of human. LIFN is a kind of spiking neuron. In LIFN, Membrane potential between spikes follows a linear differential equation. At threshold, the neuron fires a spike, and the membrane potential is reset to resting membrane potential. If a group of spikes is considered a pulse train, it is possible to introduce LIFN into CNN(Cellular Neural Network) instead of ΣΔmodulation. In this manuscript, Performance on image restoration of LIFN-CNN is compared with that of ΣΔ-CNN by calculating PSNR(Peak Signal to Noise Ratio).

The sigma-delta cellular neural network (SD-CNN) is a novel framework of spatial domain sigma-delta modulation utilizing neuro dynamics. Also, it has signal reconstruction and noise shaping characteristics that are important sigma-delta properties. Although the noise shaping effect with the oversampling technique plays very important role for drastic quantization noise reduction in binary digital sequences, the conventional SD-CNN could not use it effectively due to the fact that an oversampling factor of the SD-CNN has been undissolved mathematically. In this paper, a novel SD-CNN with the oversampling technique having second order noise shaping property by multi-stage noise shaping (MASH) is proposed. Moreover, the oversampling factor of SD-CNN is demonstrated mathematically. Experimental results of various standard test images in several oversampling ratios suggest that the proposed oversampling SD-CNN has an excellent AD and DA performance.

In this paper, a novel lossless image coding method based on the lifting wavelet transform using discrete-time cellular neural networks (DT-CNNs) with chaos noise is proposed. In our method, the image is interpolated by using the nonlinear interpolative dynamics of DTCNNs. Moreover, the adaptive chaos noise is used to avoid the local minimimum problem of DT-CNNs dynamics for the improvement of the prediction ability of DT-CNNs. The experimental results shows a better coding performance compared with the conventional methods.

In our previous research, the Maximum-Flow Neural Network (MF-NN) was proposed, and we showed that the MF-NN is possible to solve any maximum-flow problems. For application to the maximum-flow algorithm, the sigmoidal function f(x) is applied as a nonlinear function having saturation characteristic. However, the sigmoidal function never converges f(x)=0,1 which is vital values as the maximum-flow algorithm. In this research, we propose novel MF-NN using piecewise linear function for improving those problems. Moreover, this novel method is possible to considerably reduce a calculation cost.

In this paper, we propose a method of orbit control of nonlinear manipulators using path search by local current comparison method. The local current comparison method is a effective method of solving shortest path problem. By using this method, the shortest orbit that can avoid obstacles is generated automatically.

The sigma-delta cellular neural network (SD-CNN) is a complete framework of a spatial domain sigma-delta modulator, and has a very high image reconstruction (AD-to-DA) performance. In this architecture, the A-template given by a 2-D low pass filter (LPF) is used for a digital to analogue converter (DAC), the C-template works as an integrator, and the nonlinear output function is for the bilevel output. By exploiting to the nonlinear optimization ability of CNN spatio-temporal dynamics, optimal binary and reconstruction image can be obtained. However, in the conventional SD-CNN, the Gaussian LPF, whose coefficients are real number, is used as the A-template. This filter coefficients requirement is one of major factors that restricts a hardware implementation. In this paper, to deal with abovementioned limitation difficulty, a standard 2-D LPF is used for the A-template, and the effectiveness of the proposed method is confirmed by the experimental results.

A sigma-delta modulation is a well-known concept for analog-to-digital (A/D) converter. However, its underlying concept is limited to 1-D signals. The Sigma-Delta Cellular Neural Network (SD-CNN) is an efficient framework for a spatial domain sigma-delta modulation. Due to a CNN dynamics, each pixel of an image corresponds to a cell of a CNN, and each cell is connected spatially by the A-template. Therefore, the SD-CNN can be thought of as a very large-scale and super-parallel sigma-delta modulator. In this paper, we propose a novel functional sigma-delta modulation using SD-CNN. In order to provide new functions for SD-CNN, the essential conditions for constructing a spatial-domain sigma-delta modulation of CNN are reexamined. Multibit SD-CNN for high image reconstruction performance and SD-CNN with basic image processing ability for an integrated camera interface are proposed in this paper.

In this paper, image resolution compression based on retina model using DT-CNN is proposed. By using sigma-delta modulator by CNN, the input image can be converted into digital pulse sequences, and the image can be reconstructed. Human has 100 million retinal cells or more, and the input signal via the retina is sent to the cerebrum visual field. The signal from there is transmitted to the cerebrum visual field through the optic nerve fiber of about one million. In a word, the resolution of input image is compressed, and converted into binary digital purse sequences in the system from the retina to the cerebrum visual field. These binary digital pulse sequences are sent to the cerebrum visual field. The transmitted binary digital pulse sequences are reconstructed in the brain finally. That is, a model from the retina to the cerebrum can be achieved by using CNN. The experimental results show that a good quality reconstruction resolution compressed image was able to be obtained, and the image resolution compression by CNN based on the retina model was able to be achieved.

The technique for adding images of binary is the best for the step of the image interpolation. Because making binary image by the output function with two saturation area is possible, CNN can be applied to Digetal harftorning by designing the paramata and template of discrete time cellular neural network (DT-CNN). We pays attention to a similar sigma delta modulation as a system, and use nonlinear interpolative effect of its dyanmics to obtain the step interpolation imagefrom images of binary. In this paper, we verifies it by experimenting on the step interpolation performance of the proposal method to various images by using the space area sigma delta modulation characteristic of CNN.

In this paper, we propose a new numerical analysis method for data mining based on covariance structure. We use Backward Euler method to minimize the difference between the covariance matrix of the model and the covariance matrix of the measurement data. Then the quasi-Newton method is used to modify the next step solution. Experimental results show that the performance of our proposed method has been better than that of conventional methods. In addition, it could be applied to pruning of passes in the covariance structure model.

Recently, collecting a lot of data has become easy with advancement of information technology. Automatic classification system is required because it is very difficult that a lot of data is sorted out for analysis. In order to classify large-scale data set automatically, many aigorithms require classes for records respectively in learning process. However, the investigation of classes included in data set is very expensive and it takes a long time. Then, a lot of research for automatically cluster generation algorithms are carried out up to now. But, many algorithms require the number of clusters and adjustments of parameters for each problem. We propose an automatically cluster generation algorithm without the number of clusters and adjustments of parameters for each problem by dynamic transition decision trees changing from initial state to final state. It is proved base on a new entropy theory for bottom up decision tree that the dynamics can converges to almost optimal equilibrium point of tree structure.

The lifting scheme is a flexible method for the construction of linear and nonlinear wavelet transforms. In this paper, we propose a novel lossless high dynamic range (HDR) image coding method based on the lifting scheme using discrete-time cellular neural networks (DT-CNNs). In our proposed method, the image is interpolated by using the nonlinear interpolative dynamics of DT-CNN. Since the output function of DT-CNN works as a multi-level quantization function, our method adapts for the prediction of HDR image, and composes the integer lifting scheme for lossless coding. Moreover, our method makes good use of the nonlinear interpolative dynamics by A-template compared with conventional CNN image coding methods using only B-template. The experimental results show a better coding performance compared with the conventional lifting method using linear filters.

This paper presents a novel class of cellular neural network (CNN), where output of a cell in the CNN is given by the piecewise linear function having multiple constant regions or quantization function. CNN with one of these output functions allow us to extend CNN to image processing with multiple glay levels. Since each cell of the original CNN has the piecewise linear output function with the two saturation regions, the image processing tasks afe mainly developed for a black and white output image. Hence, the proposed architecture will extend the promising nature of CNN further. Moreover, the hysteresis characteristics are introduced for these functions, which make tolerance to a noise robust. It is demonstrated mathematically that under a mild assumption the stability of the CNN which has output function with hysteresis characteristics is guaranteed.

Adding the dimension of time to databases produces time series databases (TSDB) and introduces new aspects and difficulties for data mining, knowledge discovery and prediction of sample points. In this paper, we introduce the method for the prediction of the next sample point with multi-layer neural network in TSDB. Predicting the next sample point in TSDB includes cleaning and filtering the time series data, identifying the most important predicting attributes, and extracting a set of association rules that can be used to predict the time series behavior in the future. Our method is based on signal processing techniques, and TSDB for the closing price of a stock are used as an example.

We proposed previously the cyclic type of sigma delta ADC that a feedback line is equipped with a conventional sigma delta ADC for improving resolution and reducing conversion time. Also the TDM (Time Division Multiplexing) cyclic sigma delta ADC architecture was proposed on previous research. In this paper, the circuit of TDM cyclic sigma delta ADC is designed. And the circuit is confirmed with major circuit simulator. To achieve the circuit of TDM cyclic sigma delta ADC,we introduced new algorithm of the Inverse Integrator output. This algorithm has a little difference of usual sigma delta schema.

In this report, the passivity of the linear cellular neural network is discussed. Consequently, it is showed that the stability condition provided by the previous works is not correct and the passivity is a part of the stable region. From the fact, we point out a necessity for reconsideration of the stability of the cellular neural networks.

Hysteresis neural networks are one of artificial neural networks which work effectively against large scale problems. In the previous work, remarkable methods have never been developed to overcome the defects of hysteresis neural networks. We then propose a novel architecture for combinatorial optimization problems to overcome them. Experimental results indicate the efficiency of the architecture.

Wavelet transform is one of well known method of digital image processing. In a practical image processing, wavelet transform requires the function orthogonality for reconstruction of the original image. The orthogonality has disadvantage for selected filter. However, it is not necessary to select an orthogonal templates in Cellular Neural Network(CNN)image processing, because the CNN is nonlinear analog circuit to obtain equilibrium points automatically and simultaneously. So, we propose analog non-orthogonal wavelet transform by CNN.

The Sparse Gaussian Radial Basis Function network for image interpolation is proposed.The center of the Gaussian Radial Basis Function is calculated by a relaxation method, whose computational process is the same with the dynamics of DT-CNN.The DT-CNN is easily realized by digital hardware technology.Thus, the proposed coding system by the Sparse Gaussian Radial Basis Function network is suitable for hardware implementation.

Hysteresis neural network is applied into combinatorial optimization problem and efficiency of its parallel computing is obtained by numerical calculations.In this research, we implement hardware optimization problems solver by hysteresis neural networks.To produce hysteresis neural module, we propose a novel synapse architecture.From experimental results, we confirm the efficiency of implementation.

In image interpolation by Radial Basis Function (RBF) Networks, using Kronecker product were proposed, in order to reduce the computational time and computer memory. However, it is only applicable to square image has same vertical and horizontal size. In this paper, the method is generalized so that it can be applied to rectangular image.

Mocromodeling of subnetwork characterized by sampled data is very important for microwave and circuit mixed mode simulation. In this case, the sampled data in the frequency-domain must be approximated by rational function. Here, this paper discusses extension of adaptive least square method and presents accurate and efficient rational approximation method.

If a digital image with minimum information quantity can be generated from a beautiful analog original real image and an original digital image equivalent with the analog original image can be generated, we call it a semi-reversible transmission. This paper describes a cellular neural network (CNN) which is used to the semi-reversible transmission. The semi-reversible transmission can be realized only by use of the CNN with a nonlinear interpolation effect generated by CNN dynamics.

A new order-reduction method of interconnect networks is presented. In this method, the impulse response of interconnect networks in the frequency-domain is approximated by rational function complexs, and a set of poles is selected by orthogonal least square algorithm, following dependency of a pole to the impulse response in the frequency-domain. Our method is applicable to the analysis of interconnect networks with measured or calculated data.

In this paper, we propose a method for solving the inverse kinematics problem with black box nonlinearity, where it is assumed that the control angles and end effector as respectively input and output signals are obtained. This method finds the solutions by using pseudo-inverse Jacobian matrix, and the Jacobian matrix is derived from the displacement in obtaining a small control signal. In computer simulation, we confirmed that the proposed method can be found the solution of the inverse kinematics problem.

The cooperative and competitive networks proposed by Amari and Abib are mathmatical model only. If this networks are implemented on realistic circuit elements, there are many problem to do. In this paper, the Amari-Abib model is extended so that it is easy to realize by circuit elements. Moreover, the extended cooperative and competitive networks are applied to block matching of a image, in order to illustrate that the proposed model is more effective.

We proposes hysteresis neural networks for combinatorial optimaization problems. In this article, we treat "N-Queen Problems" in the combinatorial optimization problems. Tank and Hopfield have proposed a linear programming solver with a neural network. The network can seek a minimum of an energy function, but they did not prove that this minimum corresponds to the solution of the problems. Our system does not define the monotone decreasing energy function, then the system may have an oscillating state. However, this system guantees that a stable equilibrium point corresponds to a solution of the combinatorial optimization problems. So, this system can solve the N-queen problems efficiently.

This article consider an optimaization problems solver using a neural network. Tank and Hopfield have proposed a linear programming solver with a neural network. The network can seek a minimum of an energy function, but they did not prove that this minimum corresponds to the solution of the problems. In this article, we popopose a novel synthesis procedure which can seek a minimum of a cost function for an optimization problems. Our system does not define the energy function, then the system may have an oscillating state. However, this system guantees that a fixed point corresponds to a minimum of a cost function.

This paper describes problems on autoresolution by Hierarchical Discrete-Time Cellular Neural Network (DT-CNN). Quantized density of the secondary color can be generated from state variables in form of memoryless by using local DT-CNN dynamics of subpixels in each pixel. Optimal pixel density corresponding to a resolution can be controlled by selection of both the order and the number of subpixels in each pixel.

We propose a new method for the local dynamics of hierarchical continuous time cellular neural network (CT-CNN). The local dynamics with different connection coefficients in each subpixels converts an analog input signal into a suitable quantization value by determining suitable different time constant.