Ikuo Matsuba

Application of neural sequential associator to long-term stock price prediction

A neural sequential associator using feedback multilayer neural networks in duplicate is proposed to analyze the inherent structure in the sequence and to predict the future sequence based on this structure. It is shown that the present method gives a better performance than that of neural networks without feedback when applied to the prediction of long-term stock prices.<<ETX>>

Renormalization group approach to hierarchical image analysis

Image analysis is the most important step of image understanding. Changes in spatial image structure must be detected at different levels of detail and over different extents in order to extract image features at different scales from noisy images. The author formulates the problem as the minimization of an image energy that combines a smoothness term, a discrepancy term, and a nonlinear term. A hierarchical method for image analysis is established by applying the renormalization group procedure to the image energy. Simulation shows that the well-segmented images are obtained hierarchically, and that this approach is useful for coarse-to-fine matching in image analysis.<<ETX>>

Neural sequential associator and its application to stock price prediction

A neural sequential associator using feedback multilayer neural networks is proposed to predict long-term time series data. The neural network analyzes the inherent structure in the sequence and predicts the future sequence based on these structures. Feedback multilayer neural networks are used in duplicate and the inputs to such models are functions of time to represent time correlations of temporal data in the synaptic weights during learning. It is shown that the method gives better performance than neural networks without feedback when applied to the prediction of long-term stock prices.<<ETX>>

Optimizing multilayer neural networks using fractal dimensions of time-series data

A fractal dimension of time-series data is used to optimize a three-layer feedback neural network which was proposed previously to detect an important time structure of time-series data and to predict a future sequence based on a current input sequence. Optimization means in a sense that the prediction error is minimized. A time interval giving the same fractal dimensions is used as an optimal size of the output layer. The number of input units is twice the number of output units. It is also found that reliability in prediction is determined empirically as a function of the fractal dimension.<<ETX>>

Chapter 8 Processing and Semiconductor Thermoelastic Behavior

Publisher Summary This chapter discusses processing and semiconductor thermoelastic behavior. For many years, the standard technique has been used for annealing and growing silicon oxide films in the fabrication of large-scale integrated circuits. The silicon wafers are held vertically on a vitreous silica boat designed so that they sit in approximately an axially symmetric position in the silica tube of the diffusion furnace. Besides cost and reliability, a major advantage of this technique over other methods is the uniformity of film thickness within a wafer and from wafer to wafer. The two most interesting variables are defects and film thickness, which depend on a number of process variables. The most important ones are temperature distribution in wafers and inlet gas partial pressure. It is a good approximation that a convective layer of gas ends at the surface of the cylinder that envelops the wafers and that the gas is stagnant among them because the space among the wafers is very small compared with the wafer diameter.

Thermoelastic model of dislocations in wafers

The temperature and thermal stress distributions in regularly spaced circular wafers in a row in a furnace which has been used for semiconductor fabrication processes are investigated both theoretically and experimentally. A mathematical thermoelastic model is proposed to estimate the transient temperature and stress distributions in wafers of various sizes and of various row positions. The process conditions necessary to prevent defects (dislocations) are established by comparing the yield stress with the calculated stress resolved on the {111} planes in the directions. The results are in good agreement with experiments.

Asymptotic behaviors of simulated annealing and mean-field approximate annealing

The asymptotic behaviors of simulated annealing are investigated both theoretically and numerically for the fully connected Hopfield neural network. The energy using ordinary simulated annealing is found to scale as 1/log(t), while the energy using the mean-field approximate annealing method proposed in the present work shows a faster scaling property described by 1/t/sup 1/2/, where t is an iterative time step.<<ETX>>

Inherent structure detection by neural sequential associator

A sequential associator based on a feedback multilayer neural network is proposed to analyze inherent structures in a sequence generated by a nonlinear dynamical system and to predict a future sequence based on these structures. The network represents time correlations in the connection weights during learning. It is capable of detecting the inherent structure and explaining the behavior of systems. The structure of the neural sequential associator, inherent structure detection, and the optimal network size based on the use of an information criterion are discussed.<<ETX>>