Takashi Kimoto

Stock market prediction system with modular neural networks

A discussion is presented of a buying- and selling-time prediction system for stocks on the Tokyo Stock Exchange and the analysis of internal representation. The system is based on modular neural networks. The authors developed a number of learning algorithms and prediction methods for the TOPIX (Tokyo Stock Exchange Prices Indexes) prediction system. The prediction system achieved accurate predictions, and the simulation on stocks trading showed an excellent profit

Multi-sensor fusion model for constructing internal representation using autoencoder neural networks

In this paper, we propose a multi-sensor fusion model using an autoencoder neural network for 3D object recognition, which fuses multiple sensory data to integrate its internal object representation. This model was evaluated using camera images from many viewpoints on a hemisphere around the target. Three images were generated from one camera image by hue and saturation value clusters. After learning the targets images from many viewpoints in an autoencoder neural network, the continuous internal representations which correspond to viewpoints, were constructed in a compress layer of the autoencoder neural network. We found that the internal representation is generalized about the viewpoints which were not in the training sets of the target. The average of the squared errors of the autoencoder neural network is about three times higher when the compared object is unknown than when the object has already been taught as the target but not the learning point. Results of the experiment demonstrate the effectiveness of our proposed model to 3D object recognition.

A sensory information processing system using neural networks

In order to carry out actions particular to the goals, a robot processes sensory information, that is, it transforms sensed data to internal representation. In some cases, the robots internal representation cannot be determined uniquely from the sensed data. An architecture is proposed for a sensory information processing system that overcomes this ill-posed problem. The system uses an artificial neural network which is trained to transform internal representation to sensory data. Applying an iterative scheme to the network, the unique internal representation can be determined. The scheme compares the networks output (sensory data) with the sensed data, and by backpropagating the difference through the layers updates an input (internal representation) which could have created the applied output (sensed data) based on the gradient descent method. By predicting the resulting state based on the intention of the systems own movement, the accuracy and speed of sensory information processing can be improved. Simulation results for three-dimensional object recognition are given.<<ETX>>

Hierarchical sensory-motor fusion model with neural networks

Human beings recognize the physical world by integrating a variety of sensory inputs, information acquired by their own actions, and their knowledge of the world using a hierarchical parallel distributed mechanism. Sensor fusion technology focuses on imitating this mechanism and is intended for advanced sensing systems with abilities exceeding those of unimodal sensory information processing systems. Our study of sensor fusion aims to develop a hierarchical sensory motor fusion mechanism capable of intentional sensing: the concept that sensing has a goal and sensing behavior must be oriented to achieve this goal. In this paper, we propose a hierarchical sensory motor fusion model with neural networks for intentional sensing. The model propagates intentions, tightly couples recognition and action, and can perform different tasks flexibly. The model includes an algorithm called iterative inversion, which we also propose for making use of multilayer neural networks as a way to solve inverse problems of sensory information processing. We applied the hierarchical sensory-motor fusion model to a three-dimensional object recognition system and demonstrated the effectiveness of the model by computer simulations.<<ETX>>

Inverse modeling of dynamical system-network architecture with identification network and adaptation network

The authors describe a neural network architecture enabling inverse modeling of a nonlinear dynamical system. It consists of two neural networks, a system identification network and an adaptation network. The effectiveness of the proposed network architecture is examined by applying it to a digital mobile communication adaptive equalizer. In digital mobile communication, the problem of multipath fading caused by vehicular movement becomes a nonlinear dynamical system. The proposed network architecture is able to obtain an inverse model of such transmission channels and attain equalization of signal distortions. The performance of the proposed adaptive equalizer was evaluated by computer simulation. The bit error rate was found to decrease by one-third compared to that without an equalizer.<<ETX>>

Hierarchical sensory information processing model with neural networks

We propose a hierarchical sensory information processing model which achieves sensor fusion. This model has a hierarchical structure: autonomous processing units are interconnected in layers above the sensors, which get information from the physical world, and the actuators, which act on the physical world. Each processing unit consists of three basic modules—recognition, motor, and sensory-motor fusion. This paper focuses on the sensory-motor fusion and the recognition modules, both of which use neural networks. To demonstrate the effectiveness of the proposed processing model, we show simulation results for three-dimensional object recognition and develop a visual control architecture for a two-dimensional manipulator used for the task of catching an object.

Design and implementation of a personal computer local network

With the increase in the number of Personal Computers in offices and factories, a demand has been created to connect these independent units in a network and to share not only data but also expensive peripheral equipments. On the other hand, local area network technology has become an active area of research and development during the last few years.